a016fd1198a940a79ba4c99b964e401be7163b0c
[muen/linux.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72
73 #include <asm/io.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
77 #include <asm/tlb.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
80
81 #include "internal.h"
82
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
85 #endif
86
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
91
92 struct page *mem_map;
93 EXPORT_SYMBOL(mem_map);
94 #endif
95
96 /*
97  * A number of key systems in x86 including ioremap() rely on the assumption
98  * that high_memory defines the upper bound on direct map memory, then end
99  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
100  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
101  * and ZONE_HIGHMEM.
102  */
103 void *high_memory;
104 EXPORT_SYMBOL(high_memory);
105
106 /*
107  * Randomize the address space (stacks, mmaps, brk, etc.).
108  *
109  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110  *   as ancient (libc5 based) binaries can segfault. )
111  */
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
114                                         1;
115 #else
116                                         2;
117 #endif
118
119 static int __init disable_randmaps(char *s)
120 {
121         randomize_va_space = 0;
122         return 1;
123 }
124 __setup("norandmaps", disable_randmaps);
125
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
128
129 unsigned long highest_memmap_pfn __read_mostly;
130
131 /*
132  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
133  */
134 static int __init init_zero_pfn(void)
135 {
136         zero_pfn = page_to_pfn(ZERO_PAGE(0));
137         return 0;
138 }
139 core_initcall(init_zero_pfn);
140
141
142 #if defined(SPLIT_RSS_COUNTING)
143
144 void sync_mm_rss(struct mm_struct *mm)
145 {
146         int i;
147
148         for (i = 0; i < NR_MM_COUNTERS; i++) {
149                 if (current->rss_stat.count[i]) {
150                         add_mm_counter(mm, i, current->rss_stat.count[i]);
151                         current->rss_stat.count[i] = 0;
152                 }
153         }
154         current->rss_stat.events = 0;
155 }
156
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
158 {
159         struct task_struct *task = current;
160
161         if (likely(task->mm == mm))
162                 task->rss_stat.count[member] += val;
163         else
164                 add_mm_counter(mm, member, val);
165 }
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
168
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH  (64)
171 static void check_sync_rss_stat(struct task_struct *task)
172 {
173         if (unlikely(task != current))
174                 return;
175         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176                 sync_mm_rss(task->mm);
177 }
178 #else /* SPLIT_RSS_COUNTING */
179
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
182
183 static void check_sync_rss_stat(struct task_struct *task)
184 {
185 }
186
187 #endif /* SPLIT_RSS_COUNTING */
188
189 /*
190  * Note: this doesn't free the actual pages themselves. That
191  * has been handled earlier when unmapping all the memory regions.
192  */
193 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
194                            unsigned long addr)
195 {
196         pgtable_t token = pmd_pgtable(*pmd);
197         pmd_clear(pmd);
198         pte_free_tlb(tlb, token, addr);
199         mm_dec_nr_ptes(tlb->mm);
200 }
201
202 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
203                                 unsigned long addr, unsigned long end,
204                                 unsigned long floor, unsigned long ceiling)
205 {
206         pmd_t *pmd;
207         unsigned long next;
208         unsigned long start;
209
210         start = addr;
211         pmd = pmd_offset(pud, addr);
212         do {
213                 next = pmd_addr_end(addr, end);
214                 if (pmd_none_or_clear_bad(pmd))
215                         continue;
216                 free_pte_range(tlb, pmd, addr);
217         } while (pmd++, addr = next, addr != end);
218
219         start &= PUD_MASK;
220         if (start < floor)
221                 return;
222         if (ceiling) {
223                 ceiling &= PUD_MASK;
224                 if (!ceiling)
225                         return;
226         }
227         if (end - 1 > ceiling - 1)
228                 return;
229
230         pmd = pmd_offset(pud, start);
231         pud_clear(pud);
232         pmd_free_tlb(tlb, pmd, start);
233         mm_dec_nr_pmds(tlb->mm);
234 }
235
236 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
237                                 unsigned long addr, unsigned long end,
238                                 unsigned long floor, unsigned long ceiling)
239 {
240         pud_t *pud;
241         unsigned long next;
242         unsigned long start;
243
244         start = addr;
245         pud = pud_offset(p4d, addr);
246         do {
247                 next = pud_addr_end(addr, end);
248                 if (pud_none_or_clear_bad(pud))
249                         continue;
250                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
251         } while (pud++, addr = next, addr != end);
252
253         start &= P4D_MASK;
254         if (start < floor)
255                 return;
256         if (ceiling) {
257                 ceiling &= P4D_MASK;
258                 if (!ceiling)
259                         return;
260         }
261         if (end - 1 > ceiling - 1)
262                 return;
263
264         pud = pud_offset(p4d, start);
265         p4d_clear(p4d);
266         pud_free_tlb(tlb, pud, start);
267         mm_dec_nr_puds(tlb->mm);
268 }
269
270 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
271                                 unsigned long addr, unsigned long end,
272                                 unsigned long floor, unsigned long ceiling)
273 {
274         p4d_t *p4d;
275         unsigned long next;
276         unsigned long start;
277
278         start = addr;
279         p4d = p4d_offset(pgd, addr);
280         do {
281                 next = p4d_addr_end(addr, end);
282                 if (p4d_none_or_clear_bad(p4d))
283                         continue;
284                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
285         } while (p4d++, addr = next, addr != end);
286
287         start &= PGDIR_MASK;
288         if (start < floor)
289                 return;
290         if (ceiling) {
291                 ceiling &= PGDIR_MASK;
292                 if (!ceiling)
293                         return;
294         }
295         if (end - 1 > ceiling - 1)
296                 return;
297
298         p4d = p4d_offset(pgd, start);
299         pgd_clear(pgd);
300         p4d_free_tlb(tlb, p4d, start);
301 }
302
303 /*
304  * This function frees user-level page tables of a process.
305  */
306 void free_pgd_range(struct mmu_gather *tlb,
307                         unsigned long addr, unsigned long end,
308                         unsigned long floor, unsigned long ceiling)
309 {
310         pgd_t *pgd;
311         unsigned long next;
312
313         /*
314          * The next few lines have given us lots of grief...
315          *
316          * Why are we testing PMD* at this top level?  Because often
317          * there will be no work to do at all, and we'd prefer not to
318          * go all the way down to the bottom just to discover that.
319          *
320          * Why all these "- 1"s?  Because 0 represents both the bottom
321          * of the address space and the top of it (using -1 for the
322          * top wouldn't help much: the masks would do the wrong thing).
323          * The rule is that addr 0 and floor 0 refer to the bottom of
324          * the address space, but end 0 and ceiling 0 refer to the top
325          * Comparisons need to use "end - 1" and "ceiling - 1" (though
326          * that end 0 case should be mythical).
327          *
328          * Wherever addr is brought up or ceiling brought down, we must
329          * be careful to reject "the opposite 0" before it confuses the
330          * subsequent tests.  But what about where end is brought down
331          * by PMD_SIZE below? no, end can't go down to 0 there.
332          *
333          * Whereas we round start (addr) and ceiling down, by different
334          * masks at different levels, in order to test whether a table
335          * now has no other vmas using it, so can be freed, we don't
336          * bother to round floor or end up - the tests don't need that.
337          */
338
339         addr &= PMD_MASK;
340         if (addr < floor) {
341                 addr += PMD_SIZE;
342                 if (!addr)
343                         return;
344         }
345         if (ceiling) {
346                 ceiling &= PMD_MASK;
347                 if (!ceiling)
348                         return;
349         }
350         if (end - 1 > ceiling - 1)
351                 end -= PMD_SIZE;
352         if (addr > end - 1)
353                 return;
354         /*
355          * We add page table cache pages with PAGE_SIZE,
356          * (see pte_free_tlb()), flush the tlb if we need
357          */
358         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
359         pgd = pgd_offset(tlb->mm, addr);
360         do {
361                 next = pgd_addr_end(addr, end);
362                 if (pgd_none_or_clear_bad(pgd))
363                         continue;
364                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
365         } while (pgd++, addr = next, addr != end);
366 }
367
368 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
369                 unsigned long floor, unsigned long ceiling)
370 {
371         while (vma) {
372                 struct vm_area_struct *next = vma->vm_next;
373                 unsigned long addr = vma->vm_start;
374
375                 /*
376                  * Hide vma from rmap and truncate_pagecache before freeing
377                  * pgtables
378                  */
379                 unlink_anon_vmas(vma);
380                 unlink_file_vma(vma);
381
382                 if (is_vm_hugetlb_page(vma)) {
383                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
384                                 floor, next ? next->vm_start : ceiling);
385                 } else {
386                         /*
387                          * Optimization: gather nearby vmas into one call down
388                          */
389                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
390                                && !is_vm_hugetlb_page(next)) {
391                                 vma = next;
392                                 next = vma->vm_next;
393                                 unlink_anon_vmas(vma);
394                                 unlink_file_vma(vma);
395                         }
396                         free_pgd_range(tlb, addr, vma->vm_end,
397                                 floor, next ? next->vm_start : ceiling);
398                 }
399                 vma = next;
400         }
401 }
402
403 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
404 {
405         spinlock_t *ptl;
406         pgtable_t new = pte_alloc_one(mm, address);
407         if (!new)
408                 return -ENOMEM;
409
410         /*
411          * Ensure all pte setup (eg. pte page lock and page clearing) are
412          * visible before the pte is made visible to other CPUs by being
413          * put into page tables.
414          *
415          * The other side of the story is the pointer chasing in the page
416          * table walking code (when walking the page table without locking;
417          * ie. most of the time). Fortunately, these data accesses consist
418          * of a chain of data-dependent loads, meaning most CPUs (alpha
419          * being the notable exception) will already guarantee loads are
420          * seen in-order. See the alpha page table accessors for the
421          * smp_read_barrier_depends() barriers in page table walking code.
422          */
423         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
424
425         ptl = pmd_lock(mm, pmd);
426         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
427                 mm_inc_nr_ptes(mm);
428                 pmd_populate(mm, pmd, new);
429                 new = NULL;
430         }
431         spin_unlock(ptl);
432         if (new)
433                 pte_free(mm, new);
434         return 0;
435 }
436
437 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
438 {
439         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
440         if (!new)
441                 return -ENOMEM;
442
443         smp_wmb(); /* See comment in __pte_alloc */
444
445         spin_lock(&init_mm.page_table_lock);
446         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
447                 pmd_populate_kernel(&init_mm, pmd, new);
448                 new = NULL;
449         }
450         spin_unlock(&init_mm.page_table_lock);
451         if (new)
452                 pte_free_kernel(&init_mm, new);
453         return 0;
454 }
455
456 static inline void init_rss_vec(int *rss)
457 {
458         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
459 }
460
461 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
462 {
463         int i;
464
465         if (current->mm == mm)
466                 sync_mm_rss(mm);
467         for (i = 0; i < NR_MM_COUNTERS; i++)
468                 if (rss[i])
469                         add_mm_counter(mm, i, rss[i]);
470 }
471
472 /*
473  * This function is called to print an error when a bad pte
474  * is found. For example, we might have a PFN-mapped pte in
475  * a region that doesn't allow it.
476  *
477  * The calling function must still handle the error.
478  */
479 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
480                           pte_t pte, struct page *page)
481 {
482         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
483         p4d_t *p4d = p4d_offset(pgd, addr);
484         pud_t *pud = pud_offset(p4d, addr);
485         pmd_t *pmd = pmd_offset(pud, addr);
486         struct address_space *mapping;
487         pgoff_t index;
488         static unsigned long resume;
489         static unsigned long nr_shown;
490         static unsigned long nr_unshown;
491
492         /*
493          * Allow a burst of 60 reports, then keep quiet for that minute;
494          * or allow a steady drip of one report per second.
495          */
496         if (nr_shown == 60) {
497                 if (time_before(jiffies, resume)) {
498                         nr_unshown++;
499                         return;
500                 }
501                 if (nr_unshown) {
502                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
503                                  nr_unshown);
504                         nr_unshown = 0;
505                 }
506                 nr_shown = 0;
507         }
508         if (nr_shown++ == 0)
509                 resume = jiffies + 60 * HZ;
510
511         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
512         index = linear_page_index(vma, addr);
513
514         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
515                  current->comm,
516                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
517         if (page)
518                 dump_page(page, "bad pte");
519         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
520                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
521         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
522                  vma->vm_file,
523                  vma->vm_ops ? vma->vm_ops->fault : NULL,
524                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
525                  mapping ? mapping->a_ops->readpage : NULL);
526         dump_stack();
527         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
528 }
529
530 /*
531  * vm_normal_page -- This function gets the "struct page" associated with a pte.
532  *
533  * "Special" mappings do not wish to be associated with a "struct page" (either
534  * it doesn't exist, or it exists but they don't want to touch it). In this
535  * case, NULL is returned here. "Normal" mappings do have a struct page.
536  *
537  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
538  * pte bit, in which case this function is trivial. Secondly, an architecture
539  * may not have a spare pte bit, which requires a more complicated scheme,
540  * described below.
541  *
542  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
543  * special mapping (even if there are underlying and valid "struct pages").
544  * COWed pages of a VM_PFNMAP are always normal.
545  *
546  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
547  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
548  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
549  * mapping will always honor the rule
550  *
551  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
552  *
553  * And for normal mappings this is false.
554  *
555  * This restricts such mappings to be a linear translation from virtual address
556  * to pfn. To get around this restriction, we allow arbitrary mappings so long
557  * as the vma is not a COW mapping; in that case, we know that all ptes are
558  * special (because none can have been COWed).
559  *
560  *
561  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
562  *
563  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
564  * page" backing, however the difference is that _all_ pages with a struct
565  * page (that is, those where pfn_valid is true) are refcounted and considered
566  * normal pages by the VM. The disadvantage is that pages are refcounted
567  * (which can be slower and simply not an option for some PFNMAP users). The
568  * advantage is that we don't have to follow the strict linearity rule of
569  * PFNMAP mappings in order to support COWable mappings.
570  *
571  */
572 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
573                              pte_t pte, bool with_public_device)
574 {
575         unsigned long pfn = pte_pfn(pte);
576
577         if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
578                 if (likely(!pte_special(pte)))
579                         goto check_pfn;
580                 if (vma->vm_ops && vma->vm_ops->find_special_page)
581                         return vma->vm_ops->find_special_page(vma, addr);
582                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
583                         return NULL;
584                 if (is_zero_pfn(pfn))
585                         return NULL;
586
587                 /*
588                  * Device public pages are special pages (they are ZONE_DEVICE
589                  * pages but different from persistent memory). They behave
590                  * allmost like normal pages. The difference is that they are
591                  * not on the lru and thus should never be involve with any-
592                  * thing that involve lru manipulation (mlock, numa balancing,
593                  * ...).
594                  *
595                  * This is why we still want to return NULL for such page from
596                  * vm_normal_page() so that we do not have to special case all
597                  * call site of vm_normal_page().
598                  */
599                 if (likely(pfn <= highest_memmap_pfn)) {
600                         struct page *page = pfn_to_page(pfn);
601
602                         if (is_device_public_page(page)) {
603                                 if (with_public_device)
604                                         return page;
605                                 return NULL;
606                         }
607                 }
608
609                 if (pte_devmap(pte))
610                         return NULL;
611
612                 print_bad_pte(vma, addr, pte, NULL);
613                 return NULL;
614         }
615
616         /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
617
618         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
619                 if (vma->vm_flags & VM_MIXEDMAP) {
620                         if (!pfn_valid(pfn))
621                                 return NULL;
622                         goto out;
623                 } else {
624                         unsigned long off;
625                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
626                         if (pfn == vma->vm_pgoff + off)
627                                 return NULL;
628                         if (!is_cow_mapping(vma->vm_flags))
629                                 return NULL;
630                 }
631         }
632
633         if (is_zero_pfn(pfn))
634                 return NULL;
635
636 check_pfn:
637         if (unlikely(pfn > highest_memmap_pfn)) {
638                 print_bad_pte(vma, addr, pte, NULL);
639                 return NULL;
640         }
641
642         /*
643          * NOTE! We still have PageReserved() pages in the page tables.
644          * eg. VDSO mappings can cause them to exist.
645          */
646 out:
647         return pfn_to_page(pfn);
648 }
649
650 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
651 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
652                                 pmd_t pmd)
653 {
654         unsigned long pfn = pmd_pfn(pmd);
655
656         /*
657          * There is no pmd_special() but there may be special pmds, e.g.
658          * in a direct-access (dax) mapping, so let's just replicate the
659          * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
660          */
661         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
662                 if (vma->vm_flags & VM_MIXEDMAP) {
663                         if (!pfn_valid(pfn))
664                                 return NULL;
665                         goto out;
666                 } else {
667                         unsigned long off;
668                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
669                         if (pfn == vma->vm_pgoff + off)
670                                 return NULL;
671                         if (!is_cow_mapping(vma->vm_flags))
672                                 return NULL;
673                 }
674         }
675
676         if (pmd_devmap(pmd))
677                 return NULL;
678         if (is_zero_pfn(pfn))
679                 return NULL;
680         if (unlikely(pfn > highest_memmap_pfn))
681                 return NULL;
682
683         /*
684          * NOTE! We still have PageReserved() pages in the page tables.
685          * eg. VDSO mappings can cause them to exist.
686          */
687 out:
688         return pfn_to_page(pfn);
689 }
690 #endif
691
692 /*
693  * copy one vm_area from one task to the other. Assumes the page tables
694  * already present in the new task to be cleared in the whole range
695  * covered by this vma.
696  */
697
698 static inline unsigned long
699 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
701                 unsigned long addr, int *rss)
702 {
703         unsigned long vm_flags = vma->vm_flags;
704         pte_t pte = *src_pte;
705         struct page *page;
706
707         /* pte contains position in swap or file, so copy. */
708         if (unlikely(!pte_present(pte))) {
709                 swp_entry_t entry = pte_to_swp_entry(pte);
710
711                 if (likely(!non_swap_entry(entry))) {
712                         if (swap_duplicate(entry) < 0)
713                                 return entry.val;
714
715                         /* make sure dst_mm is on swapoff's mmlist. */
716                         if (unlikely(list_empty(&dst_mm->mmlist))) {
717                                 spin_lock(&mmlist_lock);
718                                 if (list_empty(&dst_mm->mmlist))
719                                         list_add(&dst_mm->mmlist,
720                                                         &src_mm->mmlist);
721                                 spin_unlock(&mmlist_lock);
722                         }
723                         rss[MM_SWAPENTS]++;
724                 } else if (is_migration_entry(entry)) {
725                         page = migration_entry_to_page(entry);
726
727                         rss[mm_counter(page)]++;
728
729                         if (is_write_migration_entry(entry) &&
730                                         is_cow_mapping(vm_flags)) {
731                                 /*
732                                  * COW mappings require pages in both
733                                  * parent and child to be set to read.
734                                  */
735                                 make_migration_entry_read(&entry);
736                                 pte = swp_entry_to_pte(entry);
737                                 if (pte_swp_soft_dirty(*src_pte))
738                                         pte = pte_swp_mksoft_dirty(pte);
739                                 set_pte_at(src_mm, addr, src_pte, pte);
740                         }
741                 } else if (is_device_private_entry(entry)) {
742                         page = device_private_entry_to_page(entry);
743
744                         /*
745                          * Update rss count even for unaddressable pages, as
746                          * they should treated just like normal pages in this
747                          * respect.
748                          *
749                          * We will likely want to have some new rss counters
750                          * for unaddressable pages, at some point. But for now
751                          * keep things as they are.
752                          */
753                         get_page(page);
754                         rss[mm_counter(page)]++;
755                         page_dup_rmap(page, false);
756
757                         /*
758                          * We do not preserve soft-dirty information, because so
759                          * far, checkpoint/restore is the only feature that
760                          * requires that. And checkpoint/restore does not work
761                          * when a device driver is involved (you cannot easily
762                          * save and restore device driver state).
763                          */
764                         if (is_write_device_private_entry(entry) &&
765                             is_cow_mapping(vm_flags)) {
766                                 make_device_private_entry_read(&entry);
767                                 pte = swp_entry_to_pte(entry);
768                                 set_pte_at(src_mm, addr, src_pte, pte);
769                         }
770                 }
771                 goto out_set_pte;
772         }
773
774         /*
775          * If it's a COW mapping, write protect it both
776          * in the parent and the child
777          */
778         if (is_cow_mapping(vm_flags) && pte_write(pte)) {
779                 ptep_set_wrprotect(src_mm, addr, src_pte);
780                 pte = pte_wrprotect(pte);
781         }
782
783         /*
784          * If it's a shared mapping, mark it clean in
785          * the child
786          */
787         if (vm_flags & VM_SHARED)
788                 pte = pte_mkclean(pte);
789         pte = pte_mkold(pte);
790
791         page = vm_normal_page(vma, addr, pte);
792         if (page) {
793                 get_page(page);
794                 page_dup_rmap(page, false);
795                 rss[mm_counter(page)]++;
796         } else if (pte_devmap(pte)) {
797                 page = pte_page(pte);
798
799                 /*
800                  * Cache coherent device memory behave like regular page and
801                  * not like persistent memory page. For more informations see
802                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
803                  */
804                 if (is_device_public_page(page)) {
805                         get_page(page);
806                         page_dup_rmap(page, false);
807                         rss[mm_counter(page)]++;
808                 }
809         }
810
811 out_set_pte:
812         set_pte_at(dst_mm, addr, dst_pte, pte);
813         return 0;
814 }
815
816 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
817                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
818                    unsigned long addr, unsigned long end)
819 {
820         pte_t *orig_src_pte, *orig_dst_pte;
821         pte_t *src_pte, *dst_pte;
822         spinlock_t *src_ptl, *dst_ptl;
823         int progress = 0;
824         int rss[NR_MM_COUNTERS];
825         swp_entry_t entry = (swp_entry_t){0};
826
827 again:
828         init_rss_vec(rss);
829
830         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
831         if (!dst_pte)
832                 return -ENOMEM;
833         src_pte = pte_offset_map(src_pmd, addr);
834         src_ptl = pte_lockptr(src_mm, src_pmd);
835         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
836         orig_src_pte = src_pte;
837         orig_dst_pte = dst_pte;
838         arch_enter_lazy_mmu_mode();
839
840         do {
841                 /*
842                  * We are holding two locks at this point - either of them
843                  * could generate latencies in another task on another CPU.
844                  */
845                 if (progress >= 32) {
846                         progress = 0;
847                         if (need_resched() ||
848                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
849                                 break;
850                 }
851                 if (pte_none(*src_pte)) {
852                         progress++;
853                         continue;
854                 }
855                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
856                                                         vma, addr, rss);
857                 if (entry.val)
858                         break;
859                 progress += 8;
860         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
861
862         arch_leave_lazy_mmu_mode();
863         spin_unlock(src_ptl);
864         pte_unmap(orig_src_pte);
865         add_mm_rss_vec(dst_mm, rss);
866         pte_unmap_unlock(orig_dst_pte, dst_ptl);
867         cond_resched();
868
869         if (entry.val) {
870                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
871                         return -ENOMEM;
872                 progress = 0;
873         }
874         if (addr != end)
875                 goto again;
876         return 0;
877 }
878
879 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
880                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
881                 unsigned long addr, unsigned long end)
882 {
883         pmd_t *src_pmd, *dst_pmd;
884         unsigned long next;
885
886         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
887         if (!dst_pmd)
888                 return -ENOMEM;
889         src_pmd = pmd_offset(src_pud, addr);
890         do {
891                 next = pmd_addr_end(addr, end);
892                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
893                         || pmd_devmap(*src_pmd)) {
894                         int err;
895                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
896                         err = copy_huge_pmd(dst_mm, src_mm,
897                                             dst_pmd, src_pmd, addr, vma);
898                         if (err == -ENOMEM)
899                                 return -ENOMEM;
900                         if (!err)
901                                 continue;
902                         /* fall through */
903                 }
904                 if (pmd_none_or_clear_bad(src_pmd))
905                         continue;
906                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
907                                                 vma, addr, next))
908                         return -ENOMEM;
909         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
910         return 0;
911 }
912
913 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
914                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
915                 unsigned long addr, unsigned long end)
916 {
917         pud_t *src_pud, *dst_pud;
918         unsigned long next;
919
920         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
921         if (!dst_pud)
922                 return -ENOMEM;
923         src_pud = pud_offset(src_p4d, addr);
924         do {
925                 next = pud_addr_end(addr, end);
926                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
927                         int err;
928
929                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
930                         err = copy_huge_pud(dst_mm, src_mm,
931                                             dst_pud, src_pud, addr, vma);
932                         if (err == -ENOMEM)
933                                 return -ENOMEM;
934                         if (!err)
935                                 continue;
936                         /* fall through */
937                 }
938                 if (pud_none_or_clear_bad(src_pud))
939                         continue;
940                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
941                                                 vma, addr, next))
942                         return -ENOMEM;
943         } while (dst_pud++, src_pud++, addr = next, addr != end);
944         return 0;
945 }
946
947 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
948                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
949                 unsigned long addr, unsigned long end)
950 {
951         p4d_t *src_p4d, *dst_p4d;
952         unsigned long next;
953
954         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
955         if (!dst_p4d)
956                 return -ENOMEM;
957         src_p4d = p4d_offset(src_pgd, addr);
958         do {
959                 next = p4d_addr_end(addr, end);
960                 if (p4d_none_or_clear_bad(src_p4d))
961                         continue;
962                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
963                                                 vma, addr, next))
964                         return -ENOMEM;
965         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
966         return 0;
967 }
968
969 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
970                 struct vm_area_struct *vma)
971 {
972         pgd_t *src_pgd, *dst_pgd;
973         unsigned long next;
974         unsigned long addr = vma->vm_start;
975         unsigned long end = vma->vm_end;
976         unsigned long mmun_start;       /* For mmu_notifiers */
977         unsigned long mmun_end;         /* For mmu_notifiers */
978         bool is_cow;
979         int ret;
980
981         /*
982          * Don't copy ptes where a page fault will fill them correctly.
983          * Fork becomes much lighter when there are big shared or private
984          * readonly mappings. The tradeoff is that copy_page_range is more
985          * efficient than faulting.
986          */
987         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
988                         !vma->anon_vma)
989                 return 0;
990
991         if (is_vm_hugetlb_page(vma))
992                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
993
994         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
995                 /*
996                  * We do not free on error cases below as remove_vma
997                  * gets called on error from higher level routine
998                  */
999                 ret = track_pfn_copy(vma);
1000                 if (ret)
1001                         return ret;
1002         }
1003
1004         /*
1005          * We need to invalidate the secondary MMU mappings only when
1006          * there could be a permission downgrade on the ptes of the
1007          * parent mm. And a permission downgrade will only happen if
1008          * is_cow_mapping() returns true.
1009          */
1010         is_cow = is_cow_mapping(vma->vm_flags);
1011         mmun_start = addr;
1012         mmun_end   = end;
1013         if (is_cow)
1014                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1015                                                     mmun_end);
1016
1017         ret = 0;
1018         dst_pgd = pgd_offset(dst_mm, addr);
1019         src_pgd = pgd_offset(src_mm, addr);
1020         do {
1021                 next = pgd_addr_end(addr, end);
1022                 if (pgd_none_or_clear_bad(src_pgd))
1023                         continue;
1024                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1025                                             vma, addr, next))) {
1026                         ret = -ENOMEM;
1027                         break;
1028                 }
1029         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1030
1031         if (is_cow)
1032                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1033         return ret;
1034 }
1035
1036 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1037                                 struct vm_area_struct *vma, pmd_t *pmd,
1038                                 unsigned long addr, unsigned long end,
1039                                 struct zap_details *details)
1040 {
1041         struct mm_struct *mm = tlb->mm;
1042         int force_flush = 0;
1043         int rss[NR_MM_COUNTERS];
1044         spinlock_t *ptl;
1045         pte_t *start_pte;
1046         pte_t *pte;
1047         swp_entry_t entry;
1048
1049         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1050 again:
1051         init_rss_vec(rss);
1052         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1053         pte = start_pte;
1054         flush_tlb_batched_pending(mm);
1055         arch_enter_lazy_mmu_mode();
1056         do {
1057                 pte_t ptent = *pte;
1058                 if (pte_none(ptent))
1059                         continue;
1060
1061                 if (pte_present(ptent)) {
1062                         struct page *page;
1063
1064                         page = _vm_normal_page(vma, addr, ptent, true);
1065                         if (unlikely(details) && page) {
1066                                 /*
1067                                  * unmap_shared_mapping_pages() wants to
1068                                  * invalidate cache without truncating:
1069                                  * unmap shared but keep private pages.
1070                                  */
1071                                 if (details->check_mapping &&
1072                                     details->check_mapping != page_rmapping(page))
1073                                         continue;
1074                         }
1075                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1076                                                         tlb->fullmm);
1077                         tlb_remove_tlb_entry(tlb, pte, addr);
1078                         if (unlikely(!page))
1079                                 continue;
1080
1081                         if (!PageAnon(page)) {
1082                                 if (pte_dirty(ptent)) {
1083                                         force_flush = 1;
1084                                         set_page_dirty(page);
1085                                 }
1086                                 if (pte_young(ptent) &&
1087                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1088                                         mark_page_accessed(page);
1089                         }
1090                         rss[mm_counter(page)]--;
1091                         page_remove_rmap(page, false);
1092                         if (unlikely(page_mapcount(page) < 0))
1093                                 print_bad_pte(vma, addr, ptent, page);
1094                         if (unlikely(__tlb_remove_page(tlb, page))) {
1095                                 force_flush = 1;
1096                                 addr += PAGE_SIZE;
1097                                 break;
1098                         }
1099                         continue;
1100                 }
1101
1102                 entry = pte_to_swp_entry(ptent);
1103                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1104                         struct page *page = device_private_entry_to_page(entry);
1105
1106                         if (unlikely(details && details->check_mapping)) {
1107                                 /*
1108                                  * unmap_shared_mapping_pages() wants to
1109                                  * invalidate cache without truncating:
1110                                  * unmap shared but keep private pages.
1111                                  */
1112                                 if (details->check_mapping !=
1113                                     page_rmapping(page))
1114                                         continue;
1115                         }
1116
1117                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1118                         rss[mm_counter(page)]--;
1119                         page_remove_rmap(page, false);
1120                         put_page(page);
1121                         continue;
1122                 }
1123
1124                 /* If details->check_mapping, we leave swap entries. */
1125                 if (unlikely(details))
1126                         continue;
1127
1128                 entry = pte_to_swp_entry(ptent);
1129                 if (!non_swap_entry(entry))
1130                         rss[MM_SWAPENTS]--;
1131                 else if (is_migration_entry(entry)) {
1132                         struct page *page;
1133
1134                         page = migration_entry_to_page(entry);
1135                         rss[mm_counter(page)]--;
1136                 }
1137                 if (unlikely(!free_swap_and_cache(entry)))
1138                         print_bad_pte(vma, addr, ptent, NULL);
1139                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1140         } while (pte++, addr += PAGE_SIZE, addr != end);
1141
1142         add_mm_rss_vec(mm, rss);
1143         arch_leave_lazy_mmu_mode();
1144
1145         /* Do the actual TLB flush before dropping ptl */
1146         if (force_flush)
1147                 tlb_flush_mmu_tlbonly(tlb);
1148         pte_unmap_unlock(start_pte, ptl);
1149
1150         /*
1151          * If we forced a TLB flush (either due to running out of
1152          * batch buffers or because we needed to flush dirty TLB
1153          * entries before releasing the ptl), free the batched
1154          * memory too. Restart if we didn't do everything.
1155          */
1156         if (force_flush) {
1157                 force_flush = 0;
1158                 tlb_flush_mmu_free(tlb);
1159                 if (addr != end)
1160                         goto again;
1161         }
1162
1163         return addr;
1164 }
1165
1166 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1167                                 struct vm_area_struct *vma, pud_t *pud,
1168                                 unsigned long addr, unsigned long end,
1169                                 struct zap_details *details)
1170 {
1171         pmd_t *pmd;
1172         unsigned long next;
1173
1174         pmd = pmd_offset(pud, addr);
1175         do {
1176                 next = pmd_addr_end(addr, end);
1177                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1178                         if (next - addr != HPAGE_PMD_SIZE)
1179                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1180                         else if (zap_huge_pmd(tlb, vma, pmd, addr))
1181                                 goto next;
1182                         /* fall through */
1183                 }
1184                 /*
1185                  * Here there can be other concurrent MADV_DONTNEED or
1186                  * trans huge page faults running, and if the pmd is
1187                  * none or trans huge it can change under us. This is
1188                  * because MADV_DONTNEED holds the mmap_sem in read
1189                  * mode.
1190                  */
1191                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1192                         goto next;
1193                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1194 next:
1195                 cond_resched();
1196         } while (pmd++, addr = next, addr != end);
1197
1198         return addr;
1199 }
1200
1201 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1202                                 struct vm_area_struct *vma, p4d_t *p4d,
1203                                 unsigned long addr, unsigned long end,
1204                                 struct zap_details *details)
1205 {
1206         pud_t *pud;
1207         unsigned long next;
1208
1209         pud = pud_offset(p4d, addr);
1210         do {
1211                 next = pud_addr_end(addr, end);
1212                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1213                         if (next - addr != HPAGE_PUD_SIZE) {
1214                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1215                                 split_huge_pud(vma, pud, addr);
1216                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1217                                 goto next;
1218                         /* fall through */
1219                 }
1220                 if (pud_none_or_clear_bad(pud))
1221                         continue;
1222                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1223 next:
1224                 cond_resched();
1225         } while (pud++, addr = next, addr != end);
1226
1227         return addr;
1228 }
1229
1230 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1231                                 struct vm_area_struct *vma, pgd_t *pgd,
1232                                 unsigned long addr, unsigned long end,
1233                                 struct zap_details *details)
1234 {
1235         p4d_t *p4d;
1236         unsigned long next;
1237
1238         p4d = p4d_offset(pgd, addr);
1239         do {
1240                 next = p4d_addr_end(addr, end);
1241                 if (p4d_none_or_clear_bad(p4d))
1242                         continue;
1243                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1244         } while (p4d++, addr = next, addr != end);
1245
1246         return addr;
1247 }
1248
1249 void unmap_page_range(struct mmu_gather *tlb,
1250                              struct vm_area_struct *vma,
1251                              unsigned long addr, unsigned long end,
1252                              struct zap_details *details)
1253 {
1254         pgd_t *pgd;
1255         unsigned long next;
1256
1257         BUG_ON(addr >= end);
1258         tlb_start_vma(tlb, vma);
1259         pgd = pgd_offset(vma->vm_mm, addr);
1260         do {
1261                 next = pgd_addr_end(addr, end);
1262                 if (pgd_none_or_clear_bad(pgd))
1263                         continue;
1264                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1265         } while (pgd++, addr = next, addr != end);
1266         tlb_end_vma(tlb, vma);
1267 }
1268
1269
1270 static void unmap_single_vma(struct mmu_gather *tlb,
1271                 struct vm_area_struct *vma, unsigned long start_addr,
1272                 unsigned long end_addr,
1273                 struct zap_details *details)
1274 {
1275         unsigned long start = max(vma->vm_start, start_addr);
1276         unsigned long end;
1277
1278         if (start >= vma->vm_end)
1279                 return;
1280         end = min(vma->vm_end, end_addr);
1281         if (end <= vma->vm_start)
1282                 return;
1283
1284         if (vma->vm_file)
1285                 uprobe_munmap(vma, start, end);
1286
1287         if (unlikely(vma->vm_flags & VM_PFNMAP))
1288                 untrack_pfn(vma, 0, 0);
1289
1290         if (start != end) {
1291                 if (unlikely(is_vm_hugetlb_page(vma))) {
1292                         /*
1293                          * It is undesirable to test vma->vm_file as it
1294                          * should be non-null for valid hugetlb area.
1295                          * However, vm_file will be NULL in the error
1296                          * cleanup path of mmap_region. When
1297                          * hugetlbfs ->mmap method fails,
1298                          * mmap_region() nullifies vma->vm_file
1299                          * before calling this function to clean up.
1300                          * Since no pte has actually been setup, it is
1301                          * safe to do nothing in this case.
1302                          */
1303                         if (vma->vm_file) {
1304                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1305                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1306                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1307                         }
1308                 } else
1309                         unmap_page_range(tlb, vma, start, end, details);
1310         }
1311 }
1312
1313 /**
1314  * unmap_vmas - unmap a range of memory covered by a list of vma's
1315  * @tlb: address of the caller's struct mmu_gather
1316  * @vma: the starting vma
1317  * @start_addr: virtual address at which to start unmapping
1318  * @end_addr: virtual address at which to end unmapping
1319  *
1320  * Unmap all pages in the vma list.
1321  *
1322  * Only addresses between `start' and `end' will be unmapped.
1323  *
1324  * The VMA list must be sorted in ascending virtual address order.
1325  *
1326  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1327  * range after unmap_vmas() returns.  So the only responsibility here is to
1328  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1329  * drops the lock and schedules.
1330  */
1331 void unmap_vmas(struct mmu_gather *tlb,
1332                 struct vm_area_struct *vma, unsigned long start_addr,
1333                 unsigned long end_addr)
1334 {
1335         struct mm_struct *mm = vma->vm_mm;
1336
1337         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1338         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1339                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1340         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1341 }
1342
1343 /**
1344  * zap_page_range - remove user pages in a given range
1345  * @vma: vm_area_struct holding the applicable pages
1346  * @start: starting address of pages to zap
1347  * @size: number of bytes to zap
1348  *
1349  * Caller must protect the VMA list
1350  */
1351 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1352                 unsigned long size)
1353 {
1354         struct mm_struct *mm = vma->vm_mm;
1355         struct mmu_gather tlb;
1356         unsigned long end = start + size;
1357
1358         lru_add_drain();
1359         tlb_gather_mmu(&tlb, mm, start, end);
1360         update_hiwater_rss(mm);
1361         mmu_notifier_invalidate_range_start(mm, start, end);
1362         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1363                 unmap_single_vma(&tlb, vma, start, end, NULL);
1364         mmu_notifier_invalidate_range_end(mm, start, end);
1365         tlb_finish_mmu(&tlb, start, end);
1366 }
1367
1368 /**
1369  * zap_page_range_single - remove user pages in a given range
1370  * @vma: vm_area_struct holding the applicable pages
1371  * @address: starting address of pages to zap
1372  * @size: number of bytes to zap
1373  * @details: details of shared cache invalidation
1374  *
1375  * The range must fit into one VMA.
1376  */
1377 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1378                 unsigned long size, struct zap_details *details)
1379 {
1380         struct mm_struct *mm = vma->vm_mm;
1381         struct mmu_gather tlb;
1382         unsigned long end = address + size;
1383
1384         lru_add_drain();
1385         tlb_gather_mmu(&tlb, mm, address, end);
1386         update_hiwater_rss(mm);
1387         mmu_notifier_invalidate_range_start(mm, address, end);
1388         unmap_single_vma(&tlb, vma, address, end, details);
1389         mmu_notifier_invalidate_range_end(mm, address, end);
1390         tlb_finish_mmu(&tlb, address, end);
1391 }
1392
1393 /**
1394  * zap_vma_ptes - remove ptes mapping the vma
1395  * @vma: vm_area_struct holding ptes to be zapped
1396  * @address: starting address of pages to zap
1397  * @size: number of bytes to zap
1398  *
1399  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1400  *
1401  * The entire address range must be fully contained within the vma.
1402  *
1403  */
1404 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1405                 unsigned long size)
1406 {
1407         if (address < vma->vm_start || address + size > vma->vm_end ||
1408                         !(vma->vm_flags & VM_PFNMAP))
1409                 return;
1410
1411         zap_page_range_single(vma, address, size, NULL);
1412 }
1413 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1414
1415 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1416                         spinlock_t **ptl)
1417 {
1418         pgd_t *pgd;
1419         p4d_t *p4d;
1420         pud_t *pud;
1421         pmd_t *pmd;
1422
1423         pgd = pgd_offset(mm, addr);
1424         p4d = p4d_alloc(mm, pgd, addr);
1425         if (!p4d)
1426                 return NULL;
1427         pud = pud_alloc(mm, p4d, addr);
1428         if (!pud)
1429                 return NULL;
1430         pmd = pmd_alloc(mm, pud, addr);
1431         if (!pmd)
1432                 return NULL;
1433
1434         VM_BUG_ON(pmd_trans_huge(*pmd));
1435         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1436 }
1437
1438 /*
1439  * This is the old fallback for page remapping.
1440  *
1441  * For historical reasons, it only allows reserved pages. Only
1442  * old drivers should use this, and they needed to mark their
1443  * pages reserved for the old functions anyway.
1444  */
1445 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1446                         struct page *page, pgprot_t prot)
1447 {
1448         struct mm_struct *mm = vma->vm_mm;
1449         int retval;
1450         pte_t *pte;
1451         spinlock_t *ptl;
1452
1453         retval = -EINVAL;
1454         if (PageAnon(page))
1455                 goto out;
1456         retval = -ENOMEM;
1457         flush_dcache_page(page);
1458         pte = get_locked_pte(mm, addr, &ptl);
1459         if (!pte)
1460                 goto out;
1461         retval = -EBUSY;
1462         if (!pte_none(*pte))
1463                 goto out_unlock;
1464
1465         /* Ok, finally just insert the thing.. */
1466         get_page(page);
1467         inc_mm_counter_fast(mm, mm_counter_file(page));
1468         page_add_file_rmap(page, false);
1469         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1470
1471         retval = 0;
1472         pte_unmap_unlock(pte, ptl);
1473         return retval;
1474 out_unlock:
1475         pte_unmap_unlock(pte, ptl);
1476 out:
1477         return retval;
1478 }
1479
1480 /**
1481  * vm_insert_page - insert single page into user vma
1482  * @vma: user vma to map to
1483  * @addr: target user address of this page
1484  * @page: source kernel page
1485  *
1486  * This allows drivers to insert individual pages they've allocated
1487  * into a user vma.
1488  *
1489  * The page has to be a nice clean _individual_ kernel allocation.
1490  * If you allocate a compound page, you need to have marked it as
1491  * such (__GFP_COMP), or manually just split the page up yourself
1492  * (see split_page()).
1493  *
1494  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1495  * took an arbitrary page protection parameter. This doesn't allow
1496  * that. Your vma protection will have to be set up correctly, which
1497  * means that if you want a shared writable mapping, you'd better
1498  * ask for a shared writable mapping!
1499  *
1500  * The page does not need to be reserved.
1501  *
1502  * Usually this function is called from f_op->mmap() handler
1503  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1504  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1505  * function from other places, for example from page-fault handler.
1506  */
1507 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1508                         struct page *page)
1509 {
1510         if (addr < vma->vm_start || addr >= vma->vm_end)
1511                 return -EFAULT;
1512         if (!page_count(page))
1513                 return -EINVAL;
1514         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1515                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1516                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1517                 vma->vm_flags |= VM_MIXEDMAP;
1518         }
1519         return insert_page(vma, addr, page, vma->vm_page_prot);
1520 }
1521 EXPORT_SYMBOL(vm_insert_page);
1522
1523 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1524                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1525 {
1526         struct mm_struct *mm = vma->vm_mm;
1527         int retval;
1528         pte_t *pte, entry;
1529         spinlock_t *ptl;
1530
1531         retval = -ENOMEM;
1532         pte = get_locked_pte(mm, addr, &ptl);
1533         if (!pte)
1534                 goto out;
1535         retval = -EBUSY;
1536         if (!pte_none(*pte)) {
1537                 if (mkwrite) {
1538                         /*
1539                          * For read faults on private mappings the PFN passed
1540                          * in may not match the PFN we have mapped if the
1541                          * mapped PFN is a writeable COW page.  In the mkwrite
1542                          * case we are creating a writable PTE for a shared
1543                          * mapping and we expect the PFNs to match.
1544                          */
1545                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1546                                 goto out_unlock;
1547                         entry = *pte;
1548                         goto out_mkwrite;
1549                 } else
1550                         goto out_unlock;
1551         }
1552
1553         /* Ok, finally just insert the thing.. */
1554         if (pfn_t_devmap(pfn))
1555                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1556         else
1557                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1558
1559 out_mkwrite:
1560         if (mkwrite) {
1561                 entry = pte_mkyoung(entry);
1562                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1563         }
1564
1565         set_pte_at(mm, addr, pte, entry);
1566         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1567
1568         retval = 0;
1569 out_unlock:
1570         pte_unmap_unlock(pte, ptl);
1571 out:
1572         return retval;
1573 }
1574
1575 /**
1576  * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1577  * @vma: user vma to map to
1578  * @addr: target user address of this page
1579  * @pfn: source kernel pfn
1580  * @pgprot: pgprot flags for the inserted page
1581  *
1582  * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1583  * to override pgprot on a per-page basis.
1584  *
1585  * This only makes sense for IO mappings, and it makes no sense for
1586  * COW mappings.  In general, using multiple vmas is preferable;
1587  * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1588  * impractical.
1589  *
1590  * Context: Process context.  May allocate using %GFP_KERNEL.
1591  * Return: vm_fault_t value.
1592  */
1593 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1594                         unsigned long pfn, pgprot_t pgprot)
1595 {
1596         int err;
1597
1598         /*
1599          * Technically, architectures with pte_special can avoid all these
1600          * restrictions (same for remap_pfn_range).  However we would like
1601          * consistency in testing and feature parity among all, so we should
1602          * try to keep these invariants in place for everybody.
1603          */
1604         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1605         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1606                                                 (VM_PFNMAP|VM_MIXEDMAP));
1607         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1608         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1609
1610         if (addr < vma->vm_start || addr >= vma->vm_end)
1611                 return VM_FAULT_SIGBUS;
1612
1613         if (!pfn_modify_allowed(pfn, pgprot))
1614                 return VM_FAULT_SIGBUS;
1615
1616         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1617
1618         err = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1619                         false);
1620
1621         if (err == -ENOMEM)
1622                 return VM_FAULT_OOM;
1623         if (err < 0 && err != -EBUSY)
1624                 return VM_FAULT_SIGBUS;
1625
1626         return VM_FAULT_NOPAGE;
1627 }
1628 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1629
1630 /**
1631  * vmf_insert_pfn - insert single pfn into user vma
1632  * @vma: user vma to map to
1633  * @addr: target user address of this page
1634  * @pfn: source kernel pfn
1635  *
1636  * Similar to vm_insert_page, this allows drivers to insert individual pages
1637  * they've allocated into a user vma. Same comments apply.
1638  *
1639  * This function should only be called from a vm_ops->fault handler, and
1640  * in that case the handler should return the result of this function.
1641  *
1642  * vma cannot be a COW mapping.
1643  *
1644  * As this is called only for pages that do not currently exist, we
1645  * do not need to flush old virtual caches or the TLB.
1646  *
1647  * Context: Process context.  May allocate using %GFP_KERNEL.
1648  * Return: vm_fault_t value.
1649  */
1650 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1651                         unsigned long pfn)
1652 {
1653         return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1654 }
1655 EXPORT_SYMBOL(vmf_insert_pfn);
1656
1657 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1658 {
1659         /* these checks mirror the abort conditions in vm_normal_page */
1660         if (vma->vm_flags & VM_MIXEDMAP)
1661                 return true;
1662         if (pfn_t_devmap(pfn))
1663                 return true;
1664         if (pfn_t_special(pfn))
1665                 return true;
1666         if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1667                 return true;
1668         return false;
1669 }
1670
1671 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1672                 unsigned long addr, pfn_t pfn, bool mkwrite)
1673 {
1674         pgprot_t pgprot = vma->vm_page_prot;
1675         int err;
1676
1677         BUG_ON(!vm_mixed_ok(vma, pfn));
1678
1679         if (addr < vma->vm_start || addr >= vma->vm_end)
1680                 return VM_FAULT_SIGBUS;
1681
1682         track_pfn_insert(vma, &pgprot, pfn);
1683
1684         if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1685                 return VM_FAULT_SIGBUS;
1686
1687         /*
1688          * If we don't have pte special, then we have to use the pfn_valid()
1689          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1690          * refcount the page if pfn_valid is true (hence insert_page rather
1691          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1692          * without pte special, it would there be refcounted as a normal page.
1693          */
1694         if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1695             !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1696                 struct page *page;
1697
1698                 /*
1699                  * At this point we are committed to insert_page()
1700                  * regardless of whether the caller specified flags that
1701                  * result in pfn_t_has_page() == false.
1702                  */
1703                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1704                 err = insert_page(vma, addr, page, pgprot);
1705         } else {
1706                 err = insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1707         }
1708
1709         if (err == -ENOMEM)
1710                 return VM_FAULT_OOM;
1711         if (err < 0 && err != -EBUSY)
1712                 return VM_FAULT_SIGBUS;
1713
1714         return VM_FAULT_NOPAGE;
1715 }
1716
1717 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1718                 pfn_t pfn)
1719 {
1720         return __vm_insert_mixed(vma, addr, pfn, false);
1721 }
1722 EXPORT_SYMBOL(vmf_insert_mixed);
1723
1724 /*
1725  *  If the insertion of PTE failed because someone else already added a
1726  *  different entry in the mean time, we treat that as success as we assume
1727  *  the same entry was actually inserted.
1728  */
1729 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1730                 unsigned long addr, pfn_t pfn)
1731 {
1732         return __vm_insert_mixed(vma, addr, pfn, true);
1733 }
1734 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1735
1736 /*
1737  * maps a range of physical memory into the requested pages. the old
1738  * mappings are removed. any references to nonexistent pages results
1739  * in null mappings (currently treated as "copy-on-access")
1740  */
1741 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1742                         unsigned long addr, unsigned long end,
1743                         unsigned long pfn, pgprot_t prot)
1744 {
1745         pte_t *pte;
1746         spinlock_t *ptl;
1747         int err = 0;
1748
1749         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1750         if (!pte)
1751                 return -ENOMEM;
1752         arch_enter_lazy_mmu_mode();
1753         do {
1754                 BUG_ON(!pte_none(*pte));
1755                 if (!pfn_modify_allowed(pfn, prot)) {
1756                         err = -EACCES;
1757                         break;
1758                 }
1759                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1760                 pfn++;
1761         } while (pte++, addr += PAGE_SIZE, addr != end);
1762         arch_leave_lazy_mmu_mode();
1763         pte_unmap_unlock(pte - 1, ptl);
1764         return err;
1765 }
1766
1767 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1768                         unsigned long addr, unsigned long end,
1769                         unsigned long pfn, pgprot_t prot)
1770 {
1771         pmd_t *pmd;
1772         unsigned long next;
1773         int err;
1774
1775         pfn -= addr >> PAGE_SHIFT;
1776         pmd = pmd_alloc(mm, pud, addr);
1777         if (!pmd)
1778                 return -ENOMEM;
1779         VM_BUG_ON(pmd_trans_huge(*pmd));
1780         do {
1781                 next = pmd_addr_end(addr, end);
1782                 err = remap_pte_range(mm, pmd, addr, next,
1783                                 pfn + (addr >> PAGE_SHIFT), prot);
1784                 if (err)
1785                         return err;
1786         } while (pmd++, addr = next, addr != end);
1787         return 0;
1788 }
1789
1790 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1791                         unsigned long addr, unsigned long end,
1792                         unsigned long pfn, pgprot_t prot)
1793 {
1794         pud_t *pud;
1795         unsigned long next;
1796         int err;
1797
1798         pfn -= addr >> PAGE_SHIFT;
1799         pud = pud_alloc(mm, p4d, addr);
1800         if (!pud)
1801                 return -ENOMEM;
1802         do {
1803                 next = pud_addr_end(addr, end);
1804                 err = remap_pmd_range(mm, pud, addr, next,
1805                                 pfn + (addr >> PAGE_SHIFT), prot);
1806                 if (err)
1807                         return err;
1808         } while (pud++, addr = next, addr != end);
1809         return 0;
1810 }
1811
1812 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1813                         unsigned long addr, unsigned long end,
1814                         unsigned long pfn, pgprot_t prot)
1815 {
1816         p4d_t *p4d;
1817         unsigned long next;
1818         int err;
1819
1820         pfn -= addr >> PAGE_SHIFT;
1821         p4d = p4d_alloc(mm, pgd, addr);
1822         if (!p4d)
1823                 return -ENOMEM;
1824         do {
1825                 next = p4d_addr_end(addr, end);
1826                 err = remap_pud_range(mm, p4d, addr, next,
1827                                 pfn + (addr >> PAGE_SHIFT), prot);
1828                 if (err)
1829                         return err;
1830         } while (p4d++, addr = next, addr != end);
1831         return 0;
1832 }
1833
1834 /**
1835  * remap_pfn_range - remap kernel memory to userspace
1836  * @vma: user vma to map to
1837  * @addr: target user address to start at
1838  * @pfn: physical address of kernel memory
1839  * @size: size of map area
1840  * @prot: page protection flags for this mapping
1841  *
1842  *  Note: this is only safe if the mm semaphore is held when called.
1843  */
1844 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1845                     unsigned long pfn, unsigned long size, pgprot_t prot)
1846 {
1847         pgd_t *pgd;
1848         unsigned long next;
1849         unsigned long end = addr + PAGE_ALIGN(size);
1850         struct mm_struct *mm = vma->vm_mm;
1851         unsigned long remap_pfn = pfn;
1852         int err;
1853
1854         /*
1855          * Physically remapped pages are special. Tell the
1856          * rest of the world about it:
1857          *   VM_IO tells people not to look at these pages
1858          *      (accesses can have side effects).
1859          *   VM_PFNMAP tells the core MM that the base pages are just
1860          *      raw PFN mappings, and do not have a "struct page" associated
1861          *      with them.
1862          *   VM_DONTEXPAND
1863          *      Disable vma merging and expanding with mremap().
1864          *   VM_DONTDUMP
1865          *      Omit vma from core dump, even when VM_IO turned off.
1866          *
1867          * There's a horrible special case to handle copy-on-write
1868          * behaviour that some programs depend on. We mark the "original"
1869          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1870          * See vm_normal_page() for details.
1871          */
1872         if (is_cow_mapping(vma->vm_flags)) {
1873                 if (addr != vma->vm_start || end != vma->vm_end)
1874                         return -EINVAL;
1875                 vma->vm_pgoff = pfn;
1876         }
1877
1878         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1879         if (err)
1880                 return -EINVAL;
1881
1882         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1883
1884         BUG_ON(addr >= end);
1885         pfn -= addr >> PAGE_SHIFT;
1886         pgd = pgd_offset(mm, addr);
1887         flush_cache_range(vma, addr, end);
1888         do {
1889                 next = pgd_addr_end(addr, end);
1890                 err = remap_p4d_range(mm, pgd, addr, next,
1891                                 pfn + (addr >> PAGE_SHIFT), prot);
1892                 if (err)
1893                         break;
1894         } while (pgd++, addr = next, addr != end);
1895
1896         if (err)
1897                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1898
1899         return err;
1900 }
1901 EXPORT_SYMBOL(remap_pfn_range);
1902
1903 /**
1904  * vm_iomap_memory - remap memory to userspace
1905  * @vma: user vma to map to
1906  * @start: start of area
1907  * @len: size of area
1908  *
1909  * This is a simplified io_remap_pfn_range() for common driver use. The
1910  * driver just needs to give us the physical memory range to be mapped,
1911  * we'll figure out the rest from the vma information.
1912  *
1913  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1914  * whatever write-combining details or similar.
1915  */
1916 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1917 {
1918         unsigned long vm_len, pfn, pages;
1919
1920         /* Check that the physical memory area passed in looks valid */
1921         if (start + len < start)
1922                 return -EINVAL;
1923         /*
1924          * You *really* shouldn't map things that aren't page-aligned,
1925          * but we've historically allowed it because IO memory might
1926          * just have smaller alignment.
1927          */
1928         len += start & ~PAGE_MASK;
1929         pfn = start >> PAGE_SHIFT;
1930         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1931         if (pfn + pages < pfn)
1932                 return -EINVAL;
1933
1934         /* We start the mapping 'vm_pgoff' pages into the area */
1935         if (vma->vm_pgoff > pages)
1936                 return -EINVAL;
1937         pfn += vma->vm_pgoff;
1938         pages -= vma->vm_pgoff;
1939
1940         /* Can we fit all of the mapping? */
1941         vm_len = vma->vm_end - vma->vm_start;
1942         if (vm_len >> PAGE_SHIFT > pages)
1943                 return -EINVAL;
1944
1945         /* Ok, let it rip */
1946         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1947 }
1948 EXPORT_SYMBOL(vm_iomap_memory);
1949
1950 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1951                                      unsigned long addr, unsigned long end,
1952                                      pte_fn_t fn, void *data)
1953 {
1954         pte_t *pte;
1955         int err;
1956         pgtable_t token;
1957         spinlock_t *uninitialized_var(ptl);
1958
1959         pte = (mm == &init_mm) ?
1960                 pte_alloc_kernel(pmd, addr) :
1961                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1962         if (!pte)
1963                 return -ENOMEM;
1964
1965         BUG_ON(pmd_huge(*pmd));
1966
1967         arch_enter_lazy_mmu_mode();
1968
1969         token = pmd_pgtable(*pmd);
1970
1971         do {
1972                 err = fn(pte++, token, addr, data);
1973                 if (err)
1974                         break;
1975         } while (addr += PAGE_SIZE, addr != end);
1976
1977         arch_leave_lazy_mmu_mode();
1978
1979         if (mm != &init_mm)
1980                 pte_unmap_unlock(pte-1, ptl);
1981         return err;
1982 }
1983
1984 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1985                                      unsigned long addr, unsigned long end,
1986                                      pte_fn_t fn, void *data)
1987 {
1988         pmd_t *pmd;
1989         unsigned long next;
1990         int err;
1991
1992         BUG_ON(pud_huge(*pud));
1993
1994         pmd = pmd_alloc(mm, pud, addr);
1995         if (!pmd)
1996                 return -ENOMEM;
1997         do {
1998                 next = pmd_addr_end(addr, end);
1999                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2000                 if (err)
2001                         break;
2002         } while (pmd++, addr = next, addr != end);
2003         return err;
2004 }
2005
2006 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2007                                      unsigned long addr, unsigned long end,
2008                                      pte_fn_t fn, void *data)
2009 {
2010         pud_t *pud;
2011         unsigned long next;
2012         int err;
2013
2014         pud = pud_alloc(mm, p4d, addr);
2015         if (!pud)
2016                 return -ENOMEM;
2017         do {
2018                 next = pud_addr_end(addr, end);
2019                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2020                 if (err)
2021                         break;
2022         } while (pud++, addr = next, addr != end);
2023         return err;
2024 }
2025
2026 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2027                                      unsigned long addr, unsigned long end,
2028                                      pte_fn_t fn, void *data)
2029 {
2030         p4d_t *p4d;
2031         unsigned long next;
2032         int err;
2033
2034         p4d = p4d_alloc(mm, pgd, addr);
2035         if (!p4d)
2036                 return -ENOMEM;
2037         do {
2038                 next = p4d_addr_end(addr, end);
2039                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2040                 if (err)
2041                         break;
2042         } while (p4d++, addr = next, addr != end);
2043         return err;
2044 }
2045
2046 /*
2047  * Scan a region of virtual memory, filling in page tables as necessary
2048  * and calling a provided function on each leaf page table.
2049  */
2050 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2051                         unsigned long size, pte_fn_t fn, void *data)
2052 {
2053         pgd_t *pgd;
2054         unsigned long next;
2055         unsigned long end = addr + size;
2056         int err;
2057
2058         if (WARN_ON(addr >= end))
2059                 return -EINVAL;
2060
2061         pgd = pgd_offset(mm, addr);
2062         do {
2063                 next = pgd_addr_end(addr, end);
2064                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2065                 if (err)
2066                         break;
2067         } while (pgd++, addr = next, addr != end);
2068
2069         return err;
2070 }
2071 EXPORT_SYMBOL_GPL(apply_to_page_range);
2072
2073 /*
2074  * handle_pte_fault chooses page fault handler according to an entry which was
2075  * read non-atomically.  Before making any commitment, on those architectures
2076  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2077  * parts, do_swap_page must check under lock before unmapping the pte and
2078  * proceeding (but do_wp_page is only called after already making such a check;
2079  * and do_anonymous_page can safely check later on).
2080  */
2081 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2082                                 pte_t *page_table, pte_t orig_pte)
2083 {
2084         int same = 1;
2085 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2086         if (sizeof(pte_t) > sizeof(unsigned long)) {
2087                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2088                 spin_lock(ptl);
2089                 same = pte_same(*page_table, orig_pte);
2090                 spin_unlock(ptl);
2091         }
2092 #endif
2093         pte_unmap(page_table);
2094         return same;
2095 }
2096
2097 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2098 {
2099         debug_dma_assert_idle(src);
2100
2101         /*
2102          * If the source page was a PFN mapping, we don't have
2103          * a "struct page" for it. We do a best-effort copy by
2104          * just copying from the original user address. If that
2105          * fails, we just zero-fill it. Live with it.
2106          */
2107         if (unlikely(!src)) {
2108                 void *kaddr = kmap_atomic(dst);
2109                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2110
2111                 /*
2112                  * This really shouldn't fail, because the page is there
2113                  * in the page tables. But it might just be unreadable,
2114                  * in which case we just give up and fill the result with
2115                  * zeroes.
2116                  */
2117                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2118                         clear_page(kaddr);
2119                 kunmap_atomic(kaddr);
2120                 flush_dcache_page(dst);
2121         } else
2122                 copy_user_highpage(dst, src, va, vma);
2123 }
2124
2125 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2126 {
2127         struct file *vm_file = vma->vm_file;
2128
2129         if (vm_file)
2130                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2131
2132         /*
2133          * Special mappings (e.g. VDSO) do not have any file so fake
2134          * a default GFP_KERNEL for them.
2135          */
2136         return GFP_KERNEL;
2137 }
2138
2139 /*
2140  * Notify the address space that the page is about to become writable so that
2141  * it can prohibit this or wait for the page to get into an appropriate state.
2142  *
2143  * We do this without the lock held, so that it can sleep if it needs to.
2144  */
2145 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2146 {
2147         vm_fault_t ret;
2148         struct page *page = vmf->page;
2149         unsigned int old_flags = vmf->flags;
2150
2151         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2152
2153         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2154         /* Restore original flags so that caller is not surprised */
2155         vmf->flags = old_flags;
2156         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2157                 return ret;
2158         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2159                 lock_page(page);
2160                 if (!page->mapping) {
2161                         unlock_page(page);
2162                         return 0; /* retry */
2163                 }
2164                 ret |= VM_FAULT_LOCKED;
2165         } else
2166                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2167         return ret;
2168 }
2169
2170 /*
2171  * Handle dirtying of a page in shared file mapping on a write fault.
2172  *
2173  * The function expects the page to be locked and unlocks it.
2174  */
2175 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2176                                     struct page *page)
2177 {
2178         struct address_space *mapping;
2179         bool dirtied;
2180         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2181
2182         dirtied = set_page_dirty(page);
2183         VM_BUG_ON_PAGE(PageAnon(page), page);
2184         /*
2185          * Take a local copy of the address_space - page.mapping may be zeroed
2186          * by truncate after unlock_page().   The address_space itself remains
2187          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2188          * release semantics to prevent the compiler from undoing this copying.
2189          */
2190         mapping = page_rmapping(page);
2191         unlock_page(page);
2192
2193         if ((dirtied || page_mkwrite) && mapping) {
2194                 /*
2195                  * Some device drivers do not set page.mapping
2196                  * but still dirty their pages
2197                  */
2198                 balance_dirty_pages_ratelimited(mapping);
2199         }
2200
2201         if (!page_mkwrite)
2202                 file_update_time(vma->vm_file);
2203 }
2204
2205 /*
2206  * Handle write page faults for pages that can be reused in the current vma
2207  *
2208  * This can happen either due to the mapping being with the VM_SHARED flag,
2209  * or due to us being the last reference standing to the page. In either
2210  * case, all we need to do here is to mark the page as writable and update
2211  * any related book-keeping.
2212  */
2213 static inline void wp_page_reuse(struct vm_fault *vmf)
2214         __releases(vmf->ptl)
2215 {
2216         struct vm_area_struct *vma = vmf->vma;
2217         struct page *page = vmf->page;
2218         pte_t entry;
2219         /*
2220          * Clear the pages cpupid information as the existing
2221          * information potentially belongs to a now completely
2222          * unrelated process.
2223          */
2224         if (page)
2225                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2226
2227         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2228         entry = pte_mkyoung(vmf->orig_pte);
2229         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2230         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2231                 update_mmu_cache(vma, vmf->address, vmf->pte);
2232         pte_unmap_unlock(vmf->pte, vmf->ptl);
2233 }
2234
2235 /*
2236  * Handle the case of a page which we actually need to copy to a new page.
2237  *
2238  * Called with mmap_sem locked and the old page referenced, but
2239  * without the ptl held.
2240  *
2241  * High level logic flow:
2242  *
2243  * - Allocate a page, copy the content of the old page to the new one.
2244  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2245  * - Take the PTL. If the pte changed, bail out and release the allocated page
2246  * - If the pte is still the way we remember it, update the page table and all
2247  *   relevant references. This includes dropping the reference the page-table
2248  *   held to the old page, as well as updating the rmap.
2249  * - In any case, unlock the PTL and drop the reference we took to the old page.
2250  */
2251 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2252 {
2253         struct vm_area_struct *vma = vmf->vma;
2254         struct mm_struct *mm = vma->vm_mm;
2255         struct page *old_page = vmf->page;
2256         struct page *new_page = NULL;
2257         pte_t entry;
2258         int page_copied = 0;
2259         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2260         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2261         struct mem_cgroup *memcg;
2262
2263         if (unlikely(anon_vma_prepare(vma)))
2264                 goto oom;
2265
2266         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2267                 new_page = alloc_zeroed_user_highpage_movable(vma,
2268                                                               vmf->address);
2269                 if (!new_page)
2270                         goto oom;
2271         } else {
2272                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2273                                 vmf->address);
2274                 if (!new_page)
2275                         goto oom;
2276                 cow_user_page(new_page, old_page, vmf->address, vma);
2277         }
2278
2279         if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2280                 goto oom_free_new;
2281
2282         __SetPageUptodate(new_page);
2283
2284         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2285
2286         /*
2287          * Re-check the pte - we dropped the lock
2288          */
2289         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2290         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2291                 if (old_page) {
2292                         if (!PageAnon(old_page)) {
2293                                 dec_mm_counter_fast(mm,
2294                                                 mm_counter_file(old_page));
2295                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2296                         }
2297                 } else {
2298                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2299                 }
2300                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2301                 entry = mk_pte(new_page, vma->vm_page_prot);
2302                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2303                 /*
2304                  * Clear the pte entry and flush it first, before updating the
2305                  * pte with the new entry. This will avoid a race condition
2306                  * seen in the presence of one thread doing SMC and another
2307                  * thread doing COW.
2308                  */
2309                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2310                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2311                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2312                 lru_cache_add_active_or_unevictable(new_page, vma);
2313                 /*
2314                  * We call the notify macro here because, when using secondary
2315                  * mmu page tables (such as kvm shadow page tables), we want the
2316                  * new page to be mapped directly into the secondary page table.
2317                  */
2318                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2319                 update_mmu_cache(vma, vmf->address, vmf->pte);
2320                 if (old_page) {
2321                         /*
2322                          * Only after switching the pte to the new page may
2323                          * we remove the mapcount here. Otherwise another
2324                          * process may come and find the rmap count decremented
2325                          * before the pte is switched to the new page, and
2326                          * "reuse" the old page writing into it while our pte
2327                          * here still points into it and can be read by other
2328                          * threads.
2329                          *
2330                          * The critical issue is to order this
2331                          * page_remove_rmap with the ptp_clear_flush above.
2332                          * Those stores are ordered by (if nothing else,)
2333                          * the barrier present in the atomic_add_negative
2334                          * in page_remove_rmap.
2335                          *
2336                          * Then the TLB flush in ptep_clear_flush ensures that
2337                          * no process can access the old page before the
2338                          * decremented mapcount is visible. And the old page
2339                          * cannot be reused until after the decremented
2340                          * mapcount is visible. So transitively, TLBs to
2341                          * old page will be flushed before it can be reused.
2342                          */
2343                         page_remove_rmap(old_page, false);
2344                 }
2345
2346                 /* Free the old page.. */
2347                 new_page = old_page;
2348                 page_copied = 1;
2349         } else {
2350                 mem_cgroup_cancel_charge(new_page, memcg, false);
2351         }
2352
2353         if (new_page)
2354                 put_page(new_page);
2355
2356         pte_unmap_unlock(vmf->pte, vmf->ptl);
2357         /*
2358          * No need to double call mmu_notifier->invalidate_range() callback as
2359          * the above ptep_clear_flush_notify() did already call it.
2360          */
2361         mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2362         if (old_page) {
2363                 /*
2364                  * Don't let another task, with possibly unlocked vma,
2365                  * keep the mlocked page.
2366                  */
2367                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2368                         lock_page(old_page);    /* LRU manipulation */
2369                         if (PageMlocked(old_page))
2370                                 munlock_vma_page(old_page);
2371                         unlock_page(old_page);
2372                 }
2373                 put_page(old_page);
2374         }
2375         return page_copied ? VM_FAULT_WRITE : 0;
2376 oom_free_new:
2377         put_page(new_page);
2378 oom:
2379         if (old_page)
2380                 put_page(old_page);
2381         return VM_FAULT_OOM;
2382 }
2383
2384 /**
2385  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2386  *                        writeable once the page is prepared
2387  *
2388  * @vmf: structure describing the fault
2389  *
2390  * This function handles all that is needed to finish a write page fault in a
2391  * shared mapping due to PTE being read-only once the mapped page is prepared.
2392  * It handles locking of PTE and modifying it. The function returns
2393  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2394  * lock.
2395  *
2396  * The function expects the page to be locked or other protection against
2397  * concurrent faults / writeback (such as DAX radix tree locks).
2398  */
2399 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2400 {
2401         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2402         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2403                                        &vmf->ptl);
2404         /*
2405          * We might have raced with another page fault while we released the
2406          * pte_offset_map_lock.
2407          */
2408         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2409                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2410                 return VM_FAULT_NOPAGE;
2411         }
2412         wp_page_reuse(vmf);
2413         return 0;
2414 }
2415
2416 /*
2417  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2418  * mapping
2419  */
2420 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2421 {
2422         struct vm_area_struct *vma = vmf->vma;
2423
2424         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2425                 vm_fault_t ret;
2426
2427                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2428                 vmf->flags |= FAULT_FLAG_MKWRITE;
2429                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2430                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2431                         return ret;
2432                 return finish_mkwrite_fault(vmf);
2433         }
2434         wp_page_reuse(vmf);
2435         return VM_FAULT_WRITE;
2436 }
2437
2438 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2439         __releases(vmf->ptl)
2440 {
2441         struct vm_area_struct *vma = vmf->vma;
2442
2443         get_page(vmf->page);
2444
2445         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2446                 vm_fault_t tmp;
2447
2448                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2449                 tmp = do_page_mkwrite(vmf);
2450                 if (unlikely(!tmp || (tmp &
2451                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2452                         put_page(vmf->page);
2453                         return tmp;
2454                 }
2455                 tmp = finish_mkwrite_fault(vmf);
2456                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2457                         unlock_page(vmf->page);
2458                         put_page(vmf->page);
2459                         return tmp;
2460                 }
2461         } else {
2462                 wp_page_reuse(vmf);
2463                 lock_page(vmf->page);
2464         }
2465         fault_dirty_shared_page(vma, vmf->page);
2466         put_page(vmf->page);
2467
2468         return VM_FAULT_WRITE;
2469 }
2470
2471 /*
2472  * This routine handles present pages, when users try to write
2473  * to a shared page. It is done by copying the page to a new address
2474  * and decrementing the shared-page counter for the old page.
2475  *
2476  * Note that this routine assumes that the protection checks have been
2477  * done by the caller (the low-level page fault routine in most cases).
2478  * Thus we can safely just mark it writable once we've done any necessary
2479  * COW.
2480  *
2481  * We also mark the page dirty at this point even though the page will
2482  * change only once the write actually happens. This avoids a few races,
2483  * and potentially makes it more efficient.
2484  *
2485  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2486  * but allow concurrent faults), with pte both mapped and locked.
2487  * We return with mmap_sem still held, but pte unmapped and unlocked.
2488  */
2489 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2490         __releases(vmf->ptl)
2491 {
2492         struct vm_area_struct *vma = vmf->vma;
2493
2494         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2495         if (!vmf->page) {
2496                 /*
2497                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2498                  * VM_PFNMAP VMA.
2499                  *
2500                  * We should not cow pages in a shared writeable mapping.
2501                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2502                  */
2503                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2504                                      (VM_WRITE|VM_SHARED))
2505                         return wp_pfn_shared(vmf);
2506
2507                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2508                 return wp_page_copy(vmf);
2509         }
2510
2511         /*
2512          * Take out anonymous pages first, anonymous shared vmas are
2513          * not dirty accountable.
2514          */
2515         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2516                 int total_map_swapcount;
2517                 if (!trylock_page(vmf->page)) {
2518                         get_page(vmf->page);
2519                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2520                         lock_page(vmf->page);
2521                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2522                                         vmf->address, &vmf->ptl);
2523                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2524                                 unlock_page(vmf->page);
2525                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2526                                 put_page(vmf->page);
2527                                 return 0;
2528                         }
2529                         put_page(vmf->page);
2530                 }
2531                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2532                         if (total_map_swapcount == 1) {
2533                                 /*
2534                                  * The page is all ours. Move it to
2535                                  * our anon_vma so the rmap code will
2536                                  * not search our parent or siblings.
2537                                  * Protected against the rmap code by
2538                                  * the page lock.
2539                                  */
2540                                 page_move_anon_rmap(vmf->page, vma);
2541                         }
2542                         unlock_page(vmf->page);
2543                         wp_page_reuse(vmf);
2544                         return VM_FAULT_WRITE;
2545                 }
2546                 unlock_page(vmf->page);
2547         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2548                                         (VM_WRITE|VM_SHARED))) {
2549                 return wp_page_shared(vmf);
2550         }
2551
2552         /*
2553          * Ok, we need to copy. Oh, well..
2554          */
2555         get_page(vmf->page);
2556
2557         pte_unmap_unlock(vmf->pte, vmf->ptl);
2558         return wp_page_copy(vmf);
2559 }
2560
2561 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2562                 unsigned long start_addr, unsigned long end_addr,
2563                 struct zap_details *details)
2564 {
2565         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2566 }
2567
2568 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2569                                             struct zap_details *details)
2570 {
2571         struct vm_area_struct *vma;
2572         pgoff_t vba, vea, zba, zea;
2573
2574         vma_interval_tree_foreach(vma, root,
2575                         details->first_index, details->last_index) {
2576
2577                 vba = vma->vm_pgoff;
2578                 vea = vba + vma_pages(vma) - 1;
2579                 zba = details->first_index;
2580                 if (zba < vba)
2581                         zba = vba;
2582                 zea = details->last_index;
2583                 if (zea > vea)
2584                         zea = vea;
2585
2586                 unmap_mapping_range_vma(vma,
2587                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2588                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2589                                 details);
2590         }
2591 }
2592
2593 /**
2594  * unmap_mapping_pages() - Unmap pages from processes.
2595  * @mapping: The address space containing pages to be unmapped.
2596  * @start: Index of first page to be unmapped.
2597  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2598  * @even_cows: Whether to unmap even private COWed pages.
2599  *
2600  * Unmap the pages in this address space from any userspace process which
2601  * has them mmaped.  Generally, you want to remove COWed pages as well when
2602  * a file is being truncated, but not when invalidating pages from the page
2603  * cache.
2604  */
2605 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2606                 pgoff_t nr, bool even_cows)
2607 {
2608         struct zap_details details = { };
2609
2610         details.check_mapping = even_cows ? NULL : mapping;
2611         details.first_index = start;
2612         details.last_index = start + nr - 1;
2613         if (details.last_index < details.first_index)
2614                 details.last_index = ULONG_MAX;
2615
2616         i_mmap_lock_write(mapping);
2617         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2618                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2619         i_mmap_unlock_write(mapping);
2620 }
2621
2622 /**
2623  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2624  * address_space corresponding to the specified byte range in the underlying
2625  * file.
2626  *
2627  * @mapping: the address space containing mmaps to be unmapped.
2628  * @holebegin: byte in first page to unmap, relative to the start of
2629  * the underlying file.  This will be rounded down to a PAGE_SIZE
2630  * boundary.  Note that this is different from truncate_pagecache(), which
2631  * must keep the partial page.  In contrast, we must get rid of
2632  * partial pages.
2633  * @holelen: size of prospective hole in bytes.  This will be rounded
2634  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2635  * end of the file.
2636  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2637  * but 0 when invalidating pagecache, don't throw away private data.
2638  */
2639 void unmap_mapping_range(struct address_space *mapping,
2640                 loff_t const holebegin, loff_t const holelen, int even_cows)
2641 {
2642         pgoff_t hba = holebegin >> PAGE_SHIFT;
2643         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2644
2645         /* Check for overflow. */
2646         if (sizeof(holelen) > sizeof(hlen)) {
2647                 long long holeend =
2648                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2649                 if (holeend & ~(long long)ULONG_MAX)
2650                         hlen = ULONG_MAX - hba + 1;
2651         }
2652
2653         unmap_mapping_pages(mapping, hba, hlen, even_cows);
2654 }
2655 EXPORT_SYMBOL(unmap_mapping_range);
2656
2657 /*
2658  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2659  * but allow concurrent faults), and pte mapped but not yet locked.
2660  * We return with pte unmapped and unlocked.
2661  *
2662  * We return with the mmap_sem locked or unlocked in the same cases
2663  * as does filemap_fault().
2664  */
2665 vm_fault_t do_swap_page(struct vm_fault *vmf)
2666 {
2667         struct vm_area_struct *vma = vmf->vma;
2668         struct page *page = NULL, *swapcache;
2669         struct mem_cgroup *memcg;
2670         swp_entry_t entry;
2671         pte_t pte;
2672         int locked;
2673         int exclusive = 0;
2674         vm_fault_t ret = 0;
2675
2676         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2677                 goto out;
2678
2679         entry = pte_to_swp_entry(vmf->orig_pte);
2680         if (unlikely(non_swap_entry(entry))) {
2681                 if (is_migration_entry(entry)) {
2682                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2683                                              vmf->address);
2684                 } else if (is_device_private_entry(entry)) {
2685                         /*
2686                          * For un-addressable device memory we call the pgmap
2687                          * fault handler callback. The callback must migrate
2688                          * the page back to some CPU accessible page.
2689                          */
2690                         ret = device_private_entry_fault(vma, vmf->address, entry,
2691                                                  vmf->flags, vmf->pmd);
2692                 } else if (is_hwpoison_entry(entry)) {
2693                         ret = VM_FAULT_HWPOISON;
2694                 } else {
2695                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2696                         ret = VM_FAULT_SIGBUS;
2697                 }
2698                 goto out;
2699         }
2700
2701
2702         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2703         page = lookup_swap_cache(entry, vma, vmf->address);
2704         swapcache = page;
2705
2706         if (!page) {
2707                 struct swap_info_struct *si = swp_swap_info(entry);
2708
2709                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2710                                 __swap_count(si, entry) == 1) {
2711                         /* skip swapcache */
2712                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2713                                                         vmf->address);
2714                         if (page) {
2715                                 __SetPageLocked(page);
2716                                 __SetPageSwapBacked(page);
2717                                 set_page_private(page, entry.val);
2718                                 lru_cache_add_anon(page);
2719                                 swap_readpage(page, true);
2720                         }
2721                 } else {
2722                         page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2723                                                 vmf);
2724                         swapcache = page;
2725                 }
2726
2727                 if (!page) {
2728                         /*
2729                          * Back out if somebody else faulted in this pte
2730                          * while we released the pte lock.
2731                          */
2732                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2733                                         vmf->address, &vmf->ptl);
2734                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2735                                 ret = VM_FAULT_OOM;
2736                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2737                         goto unlock;
2738                 }
2739
2740                 /* Had to read the page from swap area: Major fault */
2741                 ret = VM_FAULT_MAJOR;
2742                 count_vm_event(PGMAJFAULT);
2743                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2744         } else if (PageHWPoison(page)) {
2745                 /*
2746                  * hwpoisoned dirty swapcache pages are kept for killing
2747                  * owner processes (which may be unknown at hwpoison time)
2748                  */
2749                 ret = VM_FAULT_HWPOISON;
2750                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2751                 goto out_release;
2752         }
2753
2754         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2755
2756         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2757         if (!locked) {
2758                 ret |= VM_FAULT_RETRY;
2759                 goto out_release;
2760         }
2761
2762         /*
2763          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2764          * release the swapcache from under us.  The page pin, and pte_same
2765          * test below, are not enough to exclude that.  Even if it is still
2766          * swapcache, we need to check that the page's swap has not changed.
2767          */
2768         if (unlikely((!PageSwapCache(page) ||
2769                         page_private(page) != entry.val)) && swapcache)
2770                 goto out_page;
2771
2772         page = ksm_might_need_to_copy(page, vma, vmf->address);
2773         if (unlikely(!page)) {
2774                 ret = VM_FAULT_OOM;
2775                 page = swapcache;
2776                 goto out_page;
2777         }
2778
2779         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2780                                         &memcg, false)) {
2781                 ret = VM_FAULT_OOM;
2782                 goto out_page;
2783         }
2784
2785         /*
2786          * Back out if somebody else already faulted in this pte.
2787          */
2788         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2789                         &vmf->ptl);
2790         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2791                 goto out_nomap;
2792
2793         if (unlikely(!PageUptodate(page))) {
2794                 ret = VM_FAULT_SIGBUS;
2795                 goto out_nomap;
2796         }
2797
2798         /*
2799          * The page isn't present yet, go ahead with the fault.
2800          *
2801          * Be careful about the sequence of operations here.
2802          * To get its accounting right, reuse_swap_page() must be called
2803          * while the page is counted on swap but not yet in mapcount i.e.
2804          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2805          * must be called after the swap_free(), or it will never succeed.
2806          */
2807
2808         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2809         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2810         pte = mk_pte(page, vma->vm_page_prot);
2811         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2812                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2813                 vmf->flags &= ~FAULT_FLAG_WRITE;
2814                 ret |= VM_FAULT_WRITE;
2815                 exclusive = RMAP_EXCLUSIVE;
2816         }
2817         flush_icache_page(vma, page);
2818         if (pte_swp_soft_dirty(vmf->orig_pte))
2819                 pte = pte_mksoft_dirty(pte);
2820         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2821         arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2822         vmf->orig_pte = pte;
2823
2824         /* ksm created a completely new copy */
2825         if (unlikely(page != swapcache && swapcache)) {
2826                 page_add_new_anon_rmap(page, vma, vmf->address, false);
2827                 mem_cgroup_commit_charge(page, memcg, false, false);
2828                 lru_cache_add_active_or_unevictable(page, vma);
2829         } else {
2830                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2831                 mem_cgroup_commit_charge(page, memcg, true, false);
2832                 activate_page(page);
2833         }
2834
2835         swap_free(entry);
2836         if (mem_cgroup_swap_full(page) ||
2837             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2838                 try_to_free_swap(page);
2839         unlock_page(page);
2840         if (page != swapcache && swapcache) {
2841                 /*
2842                  * Hold the lock to avoid the swap entry to be reused
2843                  * until we take the PT lock for the pte_same() check
2844                  * (to avoid false positives from pte_same). For
2845                  * further safety release the lock after the swap_free
2846                  * so that the swap count won't change under a
2847                  * parallel locked swapcache.
2848                  */
2849                 unlock_page(swapcache);
2850                 put_page(swapcache);
2851         }
2852
2853         if (vmf->flags & FAULT_FLAG_WRITE) {
2854                 ret |= do_wp_page(vmf);
2855                 if (ret & VM_FAULT_ERROR)
2856                         ret &= VM_FAULT_ERROR;
2857                 goto out;
2858         }
2859
2860         /* No need to invalidate - it was non-present before */
2861         update_mmu_cache(vma, vmf->address, vmf->pte);
2862 unlock:
2863         pte_unmap_unlock(vmf->pte, vmf->ptl);
2864 out:
2865         return ret;
2866 out_nomap:
2867         mem_cgroup_cancel_charge(page, memcg, false);
2868         pte_unmap_unlock(vmf->pte, vmf->ptl);
2869 out_page:
2870         unlock_page(page);
2871 out_release:
2872         put_page(page);
2873         if (page != swapcache && swapcache) {
2874                 unlock_page(swapcache);
2875                 put_page(swapcache);
2876         }
2877         return ret;
2878 }
2879
2880 /*
2881  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2882  * but allow concurrent faults), and pte mapped but not yet locked.
2883  * We return with mmap_sem still held, but pte unmapped and unlocked.
2884  */
2885 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2886 {
2887         struct vm_area_struct *vma = vmf->vma;
2888         struct mem_cgroup *memcg;
2889         struct page *page;
2890         vm_fault_t ret = 0;
2891         pte_t entry;
2892
2893         /* File mapping without ->vm_ops ? */
2894         if (vma->vm_flags & VM_SHARED)
2895                 return VM_FAULT_SIGBUS;
2896
2897         /*
2898          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2899          * pte_offset_map() on pmds where a huge pmd might be created
2900          * from a different thread.
2901          *
2902          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2903          * parallel threads are excluded by other means.
2904          *
2905          * Here we only have down_read(mmap_sem).
2906          */
2907         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2908                 return VM_FAULT_OOM;
2909
2910         /* See the comment in pte_alloc_one_map() */
2911         if (unlikely(pmd_trans_unstable(vmf->pmd)))
2912                 return 0;
2913
2914         /* Use the zero-page for reads */
2915         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2916                         !mm_forbids_zeropage(vma->vm_mm)) {
2917                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2918                                                 vma->vm_page_prot));
2919                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2920                                 vmf->address, &vmf->ptl);
2921                 if (!pte_none(*vmf->pte))
2922                         goto unlock;
2923                 ret = check_stable_address_space(vma->vm_mm);
2924                 if (ret)
2925                         goto unlock;
2926                 /* Deliver the page fault to userland, check inside PT lock */
2927                 if (userfaultfd_missing(vma)) {
2928                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2929                         return handle_userfault(vmf, VM_UFFD_MISSING);
2930                 }
2931                 goto setpte;
2932         }
2933
2934         /* Allocate our own private page. */
2935         if (unlikely(anon_vma_prepare(vma)))
2936                 goto oom;
2937         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2938         if (!page)
2939                 goto oom;
2940
2941         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
2942                                         false))
2943                 goto oom_free_page;
2944
2945         /*
2946          * The memory barrier inside __SetPageUptodate makes sure that
2947          * preceeding stores to the page contents become visible before
2948          * the set_pte_at() write.
2949          */
2950         __SetPageUptodate(page);
2951
2952         entry = mk_pte(page, vma->vm_page_prot);
2953         if (vma->vm_flags & VM_WRITE)
2954                 entry = pte_mkwrite(pte_mkdirty(entry));
2955
2956         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2957                         &vmf->ptl);
2958         if (!pte_none(*vmf->pte))
2959                 goto release;
2960
2961         ret = check_stable_address_space(vma->vm_mm);
2962         if (ret)
2963                 goto release;
2964
2965         /* Deliver the page fault to userland, check inside PT lock */
2966         if (userfaultfd_missing(vma)) {
2967                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2968                 mem_cgroup_cancel_charge(page, memcg, false);
2969                 put_page(page);
2970                 return handle_userfault(vmf, VM_UFFD_MISSING);
2971         }
2972
2973         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2974         page_add_new_anon_rmap(page, vma, vmf->address, false);
2975         mem_cgroup_commit_charge(page, memcg, false, false);
2976         lru_cache_add_active_or_unevictable(page, vma);
2977 setpte:
2978         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2979
2980         /* No need to invalidate - it was non-present before */
2981         update_mmu_cache(vma, vmf->address, vmf->pte);
2982 unlock:
2983         pte_unmap_unlock(vmf->pte, vmf->ptl);
2984         return ret;
2985 release:
2986         mem_cgroup_cancel_charge(page, memcg, false);
2987         put_page(page);
2988         goto unlock;
2989 oom_free_page:
2990         put_page(page);
2991 oom:
2992         return VM_FAULT_OOM;
2993 }
2994
2995 /*
2996  * The mmap_sem must have been held on entry, and may have been
2997  * released depending on flags and vma->vm_ops->fault() return value.
2998  * See filemap_fault() and __lock_page_retry().
2999  */
3000 static vm_fault_t __do_fault(struct vm_fault *vmf)
3001 {
3002         struct vm_area_struct *vma = vmf->vma;
3003         vm_fault_t ret;
3004
3005         ret = vma->vm_ops->fault(vmf);
3006         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3007                             VM_FAULT_DONE_COW)))
3008                 return ret;
3009
3010         if (unlikely(PageHWPoison(vmf->page))) {
3011                 if (ret & VM_FAULT_LOCKED)
3012                         unlock_page(vmf->page);
3013                 put_page(vmf->page);
3014                 vmf->page = NULL;
3015                 return VM_FAULT_HWPOISON;
3016         }
3017
3018         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3019                 lock_page(vmf->page);
3020         else
3021                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3022
3023         return ret;
3024 }
3025
3026 /*
3027  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3028  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3029  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3030  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3031  */
3032 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3033 {
3034         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3035 }
3036
3037 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3038 {
3039         struct vm_area_struct *vma = vmf->vma;
3040
3041         if (!pmd_none(*vmf->pmd))
3042                 goto map_pte;
3043         if (vmf->prealloc_pte) {
3044                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3045                 if (unlikely(!pmd_none(*vmf->pmd))) {
3046                         spin_unlock(vmf->ptl);
3047                         goto map_pte;
3048                 }
3049
3050                 mm_inc_nr_ptes(vma->vm_mm);
3051                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3052                 spin_unlock(vmf->ptl);
3053                 vmf->prealloc_pte = NULL;
3054         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3055                 return VM_FAULT_OOM;
3056         }
3057 map_pte:
3058         /*
3059          * If a huge pmd materialized under us just retry later.  Use
3060          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3061          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3062          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3063          * running immediately after a huge pmd fault in a different thread of
3064          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3065          * All we have to ensure is that it is a regular pmd that we can walk
3066          * with pte_offset_map() and we can do that through an atomic read in
3067          * C, which is what pmd_trans_unstable() provides.
3068          */
3069         if (pmd_devmap_trans_unstable(vmf->pmd))
3070                 return VM_FAULT_NOPAGE;
3071
3072         /*
3073          * At this point we know that our vmf->pmd points to a page of ptes
3074          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3075          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3076          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3077          * be valid and we will re-check to make sure the vmf->pte isn't
3078          * pte_none() under vmf->ptl protection when we return to
3079          * alloc_set_pte().
3080          */
3081         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3082                         &vmf->ptl);
3083         return 0;
3084 }
3085
3086 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3087
3088 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3089 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3090                 unsigned long haddr)
3091 {
3092         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3093                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3094                 return false;
3095         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3096                 return false;
3097         return true;
3098 }
3099
3100 static void deposit_prealloc_pte(struct vm_fault *vmf)
3101 {
3102         struct vm_area_struct *vma = vmf->vma;
3103
3104         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3105         /*
3106          * We are going to consume the prealloc table,
3107          * count that as nr_ptes.
3108          */
3109         mm_inc_nr_ptes(vma->vm_mm);
3110         vmf->prealloc_pte = NULL;
3111 }
3112
3113 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3114 {
3115         struct vm_area_struct *vma = vmf->vma;
3116         bool write = vmf->flags & FAULT_FLAG_WRITE;
3117         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3118         pmd_t entry;
3119         int i;
3120         vm_fault_t ret;
3121
3122         if (!transhuge_vma_suitable(vma, haddr))
3123                 return VM_FAULT_FALLBACK;
3124
3125         ret = VM_FAULT_FALLBACK;
3126         page = compound_head(page);
3127
3128         /*
3129          * Archs like ppc64 need additonal space to store information
3130          * related to pte entry. Use the preallocated table for that.
3131          */
3132         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3133                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3134                 if (!vmf->prealloc_pte)
3135                         return VM_FAULT_OOM;
3136                 smp_wmb(); /* See comment in __pte_alloc() */
3137         }
3138
3139         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3140         if (unlikely(!pmd_none(*vmf->pmd)))
3141                 goto out;
3142
3143         for (i = 0; i < HPAGE_PMD_NR; i++)
3144                 flush_icache_page(vma, page + i);
3145
3146         entry = mk_huge_pmd(page, vma->vm_page_prot);
3147         if (write)
3148                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3149
3150         add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3151         page_add_file_rmap(page, true);
3152         /*
3153          * deposit and withdraw with pmd lock held
3154          */
3155         if (arch_needs_pgtable_deposit())
3156                 deposit_prealloc_pte(vmf);
3157
3158         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3159
3160         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3161
3162         /* fault is handled */
3163         ret = 0;
3164         count_vm_event(THP_FILE_MAPPED);
3165 out:
3166         spin_unlock(vmf->ptl);
3167         return ret;
3168 }
3169 #else
3170 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3171 {
3172         BUILD_BUG();
3173         return 0;
3174 }
3175 #endif
3176
3177 /**
3178  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3179  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3180  *
3181  * @vmf: fault environment
3182  * @memcg: memcg to charge page (only for private mappings)
3183  * @page: page to map
3184  *
3185  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3186  * return.
3187  *
3188  * Target users are page handler itself and implementations of
3189  * vm_ops->map_pages.
3190  */
3191 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3192                 struct page *page)
3193 {
3194         struct vm_area_struct *vma = vmf->vma;
3195         bool write = vmf->flags & FAULT_FLAG_WRITE;
3196         pte_t entry;
3197         vm_fault_t ret;
3198
3199         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3200                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3201                 /* THP on COW? */
3202                 VM_BUG_ON_PAGE(memcg, page);
3203
3204                 ret = do_set_pmd(vmf, page);
3205                 if (ret != VM_FAULT_FALLBACK)
3206                         return ret;
3207         }
3208
3209         if (!vmf->pte) {
3210                 ret = pte_alloc_one_map(vmf);
3211                 if (ret)
3212                         return ret;
3213         }
3214
3215         /* Re-check under ptl */
3216         if (unlikely(!pte_none(*vmf->pte)))
3217                 return VM_FAULT_NOPAGE;
3218
3219         flush_icache_page(vma, page);
3220         entry = mk_pte(page, vma->vm_page_prot);
3221         if (write)
3222                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3223         /* copy-on-write page */
3224         if (write && !(vma->vm_flags & VM_SHARED)) {
3225                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3226                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3227                 mem_cgroup_commit_charge(page, memcg, false, false);
3228                 lru_cache_add_active_or_unevictable(page, vma);
3229         } else {
3230                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3231                 page_add_file_rmap(page, false);
3232         }
3233         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3234
3235         /* no need to invalidate: a not-present page won't be cached */
3236         update_mmu_cache(vma, vmf->address, vmf->pte);
3237
3238         return 0;
3239 }
3240
3241
3242 /**
3243  * finish_fault - finish page fault once we have prepared the page to fault
3244  *
3245  * @vmf: structure describing the fault
3246  *
3247  * This function handles all that is needed to finish a page fault once the
3248  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3249  * given page, adds reverse page mapping, handles memcg charges and LRU
3250  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3251  * error.
3252  *
3253  * The function expects the page to be locked and on success it consumes a
3254  * reference of a page being mapped (for the PTE which maps it).
3255  */
3256 vm_fault_t finish_fault(struct vm_fault *vmf)
3257 {
3258         struct page *page;
3259         vm_fault_t ret = 0;
3260
3261         /* Did we COW the page? */
3262         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3263             !(vmf->vma->vm_flags & VM_SHARED))
3264                 page = vmf->cow_page;
3265         else
3266                 page = vmf->page;
3267
3268         /*
3269          * check even for read faults because we might have lost our CoWed
3270          * page
3271          */
3272         if (!(vmf->vma->vm_flags & VM_SHARED))
3273                 ret = check_stable_address_space(vmf->vma->vm_mm);
3274         if (!ret)
3275                 ret = alloc_set_pte(vmf, vmf->memcg, page);
3276         if (vmf->pte)
3277                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3278         return ret;
3279 }
3280
3281 static unsigned long fault_around_bytes __read_mostly =
3282         rounddown_pow_of_two(65536);
3283
3284 #ifdef CONFIG_DEBUG_FS
3285 static int fault_around_bytes_get(void *data, u64 *val)
3286 {
3287         *val = fault_around_bytes;
3288         return 0;
3289 }
3290
3291 /*
3292  * fault_around_bytes must be rounded down to the nearest page order as it's
3293  * what do_fault_around() expects to see.
3294  */
3295 static int fault_around_bytes_set(void *data, u64 val)
3296 {
3297         if (val / PAGE_SIZE > PTRS_PER_PTE)
3298                 return -EINVAL;
3299         if (val > PAGE_SIZE)
3300                 fault_around_bytes = rounddown_pow_of_two(val);
3301         else
3302                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3303         return 0;
3304 }
3305 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3306                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3307
3308 static int __init fault_around_debugfs(void)
3309 {
3310         void *ret;
3311
3312         ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3313                         &fault_around_bytes_fops);
3314         if (!ret)
3315                 pr_warn("Failed to create fault_around_bytes in debugfs");
3316         return 0;
3317 }
3318 late_initcall(fault_around_debugfs);
3319 #endif
3320
3321 /*
3322  * do_fault_around() tries to map few pages around the fault address. The hope
3323  * is that the pages will be needed soon and this will lower the number of
3324  * faults to handle.
3325  *
3326  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3327  * not ready to be mapped: not up-to-date, locked, etc.
3328  *
3329  * This function is called with the page table lock taken. In the split ptlock
3330  * case the page table lock only protects only those entries which belong to
3331  * the page table corresponding to the fault address.
3332  *
3333  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3334  * only once.
3335  *
3336  * fault_around_bytes defines how many bytes we'll try to map.
3337  * do_fault_around() expects it to be set to a power of two less than or equal
3338  * to PTRS_PER_PTE.
3339  *
3340  * The virtual address of the area that we map is naturally aligned to
3341  * fault_around_bytes rounded down to the machine page size
3342  * (and therefore to page order).  This way it's easier to guarantee
3343  * that we don't cross page table boundaries.
3344  */
3345 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3346 {
3347         unsigned long address = vmf->address, nr_pages, mask;
3348         pgoff_t start_pgoff = vmf->pgoff;
3349         pgoff_t end_pgoff;
3350         int off;
3351         vm_fault_t ret = 0;
3352
3353         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3354         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3355
3356         vmf->address = max(address & mask, vmf->vma->vm_start);
3357         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3358         start_pgoff -= off;
3359
3360         /*
3361          *  end_pgoff is either the end of the page table, the end of
3362          *  the vma or nr_pages from start_pgoff, depending what is nearest.
3363          */
3364         end_pgoff = start_pgoff -
3365                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3366                 PTRS_PER_PTE - 1;
3367         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3368                         start_pgoff + nr_pages - 1);
3369
3370         if (pmd_none(*vmf->pmd)) {
3371                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3372                                                   vmf->address);
3373                 if (!vmf->prealloc_pte)
3374                         goto out;
3375                 smp_wmb(); /* See comment in __pte_alloc() */
3376         }
3377
3378         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3379
3380         /* Huge page is mapped? Page fault is solved */
3381         if (pmd_trans_huge(*vmf->pmd)) {
3382                 ret = VM_FAULT_NOPAGE;
3383                 goto out;
3384         }
3385
3386         /* ->map_pages() haven't done anything useful. Cold page cache? */
3387         if (!vmf->pte)
3388                 goto out;
3389
3390         /* check if the page fault is solved */
3391         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3392         if (!pte_none(*vmf->pte))
3393                 ret = VM_FAULT_NOPAGE;
3394         pte_unmap_unlock(vmf->pte, vmf->ptl);
3395 out:
3396         vmf->address = address;
3397         vmf->pte = NULL;
3398         return ret;
3399 }
3400
3401 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3402 {
3403         struct vm_area_struct *vma = vmf->vma;
3404         vm_fault_t ret = 0;
3405
3406         /*
3407          * Let's call ->map_pages() first and use ->fault() as fallback
3408          * if page by the offset is not ready to be mapped (cold cache or
3409          * something).
3410          */
3411         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3412                 ret = do_fault_around(vmf);
3413                 if (ret)
3414                         return ret;
3415         }
3416
3417         ret = __do_fault(vmf);
3418         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3419                 return ret;
3420
3421         ret |= finish_fault(vmf);
3422         unlock_page(vmf->page);
3423         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3424                 put_page(vmf->page);
3425         return ret;
3426 }
3427
3428 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3429 {
3430         struct vm_area_struct *vma = vmf->vma;
3431         vm_fault_t ret;
3432
3433         if (unlikely(anon_vma_prepare(vma)))
3434                 return VM_FAULT_OOM;
3435
3436         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3437         if (!vmf->cow_page)
3438                 return VM_FAULT_OOM;
3439
3440         if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3441                                 &vmf->memcg, false)) {
3442                 put_page(vmf->cow_page);
3443                 return VM_FAULT_OOM;
3444         }
3445
3446         ret = __do_fault(vmf);
3447         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3448                 goto uncharge_out;
3449         if (ret & VM_FAULT_DONE_COW)
3450                 return ret;
3451
3452         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3453         __SetPageUptodate(vmf->cow_page);
3454
3455         ret |= finish_fault(vmf);
3456         unlock_page(vmf->page);
3457         put_page(vmf->page);
3458         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3459                 goto uncharge_out;
3460         return ret;
3461 uncharge_out:
3462         mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3463         put_page(vmf->cow_page);
3464         return ret;
3465 }
3466
3467 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3468 {
3469         struct vm_area_struct *vma = vmf->vma;
3470         vm_fault_t ret, tmp;
3471
3472         ret = __do_fault(vmf);
3473         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3474                 return ret;
3475
3476         /*
3477          * Check if the backing address space wants to know that the page is
3478          * about to become writable
3479          */
3480         if (vma->vm_ops->page_mkwrite) {
3481                 unlock_page(vmf->page);
3482                 tmp = do_page_mkwrite(vmf);
3483                 if (unlikely(!tmp ||
3484                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3485                         put_page(vmf->page);
3486                         return tmp;
3487                 }
3488         }
3489
3490         ret |= finish_fault(vmf);
3491         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3492                                         VM_FAULT_RETRY))) {
3493                 unlock_page(vmf->page);
3494                 put_page(vmf->page);
3495                 return ret;
3496         }
3497
3498         fault_dirty_shared_page(vma, vmf->page);
3499         return ret;
3500 }
3501
3502 /*
3503  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3504  * but allow concurrent faults).
3505  * The mmap_sem may have been released depending on flags and our
3506  * return value.  See filemap_fault() and __lock_page_or_retry().
3507  */
3508 static vm_fault_t do_fault(struct vm_fault *vmf)
3509 {
3510         struct vm_area_struct *vma = vmf->vma;
3511         vm_fault_t ret;
3512
3513         /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3514         if (!vma->vm_ops->fault)
3515                 ret = VM_FAULT_SIGBUS;
3516         else if (!(vmf->flags & FAULT_FLAG_WRITE))
3517                 ret = do_read_fault(vmf);
3518         else if (!(vma->vm_flags & VM_SHARED))
3519                 ret = do_cow_fault(vmf);
3520         else
3521                 ret = do_shared_fault(vmf);
3522
3523         /* preallocated pagetable is unused: free it */
3524         if (vmf->prealloc_pte) {
3525                 pte_free(vma->vm_mm, vmf->prealloc_pte);
3526                 vmf->prealloc_pte = NULL;
3527         }
3528         return ret;
3529 }
3530
3531 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3532                                 unsigned long addr, int page_nid,
3533                                 int *flags)
3534 {
3535         get_page(page);
3536
3537         count_vm_numa_event(NUMA_HINT_FAULTS);
3538         if (page_nid == numa_node_id()) {
3539                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3540                 *flags |= TNF_FAULT_LOCAL;
3541         }
3542
3543         return mpol_misplaced(page, vma, addr);
3544 }
3545
3546 static vm_fault_t do_numa_page(struct vm_fault *vmf)
3547 {
3548         struct vm_area_struct *vma = vmf->vma;
3549         struct page *page = NULL;
3550         int page_nid = -1;
3551         int last_cpupid;
3552         int target_nid;
3553         bool migrated = false;
3554         pte_t pte;
3555         bool was_writable = pte_savedwrite(vmf->orig_pte);
3556         int flags = 0;
3557
3558         /*
3559          * The "pte" at this point cannot be used safely without
3560          * validation through pte_unmap_same(). It's of NUMA type but
3561          * the pfn may be screwed if the read is non atomic.
3562          */
3563         vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3564         spin_lock(vmf->ptl);
3565         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3566                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3567                 goto out;
3568         }
3569
3570         /*
3571          * Make it present again, Depending on how arch implementes non
3572          * accessible ptes, some can allow access by kernel mode.
3573          */
3574         pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3575         pte = pte_modify(pte, vma->vm_page_prot);
3576         pte = pte_mkyoung(pte);
3577         if (was_writable)
3578                 pte = pte_mkwrite(pte);
3579         ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3580         update_mmu_cache(vma, vmf->address, vmf->pte);
3581
3582         page = vm_normal_page(vma, vmf->address, pte);
3583         if (!page) {
3584                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3585                 return 0;
3586         }
3587
3588         /* TODO: handle PTE-mapped THP */
3589         if (PageCompound(page)) {
3590                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3591                 return 0;
3592         }
3593
3594         /*
3595          * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3596          * much anyway since they can be in shared cache state. This misses
3597          * the case where a mapping is writable but the process never writes
3598          * to it but pte_write gets cleared during protection updates and
3599          * pte_dirty has unpredictable behaviour between PTE scan updates,
3600          * background writeback, dirty balancing and application behaviour.
3601          */
3602         if (!pte_write(pte))
3603                 flags |= TNF_NO_GROUP;
3604
3605         /*
3606          * Flag if the page is shared between multiple address spaces. This
3607          * is later used when determining whether to group tasks together
3608          */
3609         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3610                 flags |= TNF_SHARED;
3611
3612         last_cpupid = page_cpupid_last(page);
3613         page_nid = page_to_nid(page);
3614         target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3615                         &flags);
3616         pte_unmap_unlock(vmf->pte, vmf->ptl);
3617         if (target_nid == -1) {
3618                 put_page(page);
3619                 goto out;
3620         }
3621
3622         /* Migrate to the requested node */
3623         migrated = migrate_misplaced_page(page, vma, target_nid);
3624         if (migrated) {
3625                 page_nid = target_nid;
3626                 flags |= TNF_MIGRATED;
3627         } else
3628                 flags |= TNF_MIGRATE_FAIL;
3629
3630 out:
3631         if (page_nid != -1)
3632                 task_numa_fault(last_cpupid, page_nid, 1, flags);
3633         return 0;
3634 }
3635
3636 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3637 {
3638         if (vma_is_anonymous(vmf->vma))
3639                 return do_huge_pmd_anonymous_page(vmf);
3640         if (vmf->vma->vm_ops->huge_fault)
3641                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3642         return VM_FAULT_FALLBACK;
3643 }
3644
3645 /* `inline' is required to avoid gcc 4.1.2 build error */
3646 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3647 {
3648         if (vma_is_anonymous(vmf->vma))
3649                 return do_huge_pmd_wp_page(vmf, orig_pmd);
3650         if (vmf->vma->vm_ops->huge_fault)
3651                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3652
3653         /* COW handled on pte level: split pmd */
3654         VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3655         __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3656
3657         return VM_FAULT_FALLBACK;
3658 }
3659
3660 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3661 {
3662         return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3663 }
3664
3665 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3666 {
3667 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3668         /* No support for anonymous transparent PUD pages yet */
3669         if (vma_is_anonymous(vmf->vma))
3670                 return VM_FAULT_FALLBACK;
3671         if (vmf->vma->vm_ops->huge_fault)
3672                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3673 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3674         return VM_FAULT_FALLBACK;
3675 }
3676
3677 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3678 {
3679 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3680         /* No support for anonymous transparent PUD pages yet */
3681         if (vma_is_anonymous(vmf->vma))
3682                 return VM_FAULT_FALLBACK;
3683         if (vmf->vma->vm_ops->huge_fault)
3684                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3685 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3686         return VM_FAULT_FALLBACK;
3687 }
3688
3689 /*
3690  * These routines also need to handle stuff like marking pages dirty
3691  * and/or accessed for architectures that don't do it in hardware (most
3692  * RISC architectures).  The early dirtying is also good on the i386.
3693  *
3694  * There is also a hook called "update_mmu_cache()" that architectures
3695  * with external mmu caches can use to update those (ie the Sparc or
3696  * PowerPC hashed page tables that act as extended TLBs).
3697  *
3698  * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3699  * concurrent faults).
3700  *
3701  * The mmap_sem may have been released depending on flags and our return value.
3702  * See filemap_fault() and __lock_page_or_retry().
3703  */
3704 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3705 {
3706         pte_t entry;
3707
3708         if (unlikely(pmd_none(*vmf->pmd))) {
3709                 /*
3710                  * Leave __pte_alloc() until later: because vm_ops->fault may
3711                  * want to allocate huge page, and if we expose page table
3712                  * for an instant, it will be difficult to retract from
3713                  * concurrent faults and from rmap lookups.
3714                  */
3715                 vmf->pte = NULL;
3716         } else {
3717                 /* See comment in pte_alloc_one_map() */
3718                 if (pmd_devmap_trans_unstable(vmf->pmd))
3719                         return 0;
3720                 /*
3721                  * A regular pmd is established and it can't morph into a huge
3722                  * pmd from under us anymore at this point because we hold the
3723                  * mmap_sem read mode and khugepaged takes it in write mode.
3724                  * So now it's safe to run pte_offset_map().
3725                  */
3726                 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3727                 vmf->orig_pte = *vmf->pte;
3728
3729                 /*
3730                  * some architectures can have larger ptes than wordsize,
3731                  * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3732                  * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3733                  * accesses.  The code below just needs a consistent view
3734                  * for the ifs and we later double check anyway with the
3735                  * ptl lock held. So here a barrier will do.
3736                  */
3737                 barrier();
3738                 if (pte_none(vmf->orig_pte)) {
3739                         pte_unmap(vmf->pte);
3740                         vmf->pte = NULL;
3741                 }
3742         }
3743
3744         if (!vmf->pte) {
3745                 if (vma_is_anonymous(vmf->vma))
3746                         return do_anonymous_page(vmf);
3747                 else
3748                         return do_fault(vmf);
3749         }
3750
3751         if (!pte_present(vmf->orig_pte))
3752                 return do_swap_page(vmf);
3753
3754         if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3755                 return do_numa_page(vmf);
3756
3757         vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3758         spin_lock(vmf->ptl);
3759         entry = vmf->orig_pte;
3760         if (unlikely(!pte_same(*vmf->pte, entry)))
3761                 goto unlock;
3762         if (vmf->flags & FAULT_FLAG_WRITE) {
3763                 if (!pte_write(entry))
3764                         return do_wp_page(vmf);
3765                 entry = pte_mkdirty(entry);
3766         }
3767         entry = pte_mkyoung(entry);
3768         if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3769                                 vmf->flags & FAULT_FLAG_WRITE)) {
3770                 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3771         } else {
3772                 /*
3773                  * This is needed only for protection faults but the arch code
3774                  * is not yet telling us if this is a protection fault or not.
3775                  * This still avoids useless tlb flushes for .text page faults
3776                  * with threads.
3777                  */
3778                 if (vmf->flags & FAULT_FLAG_WRITE)
3779                         flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3780         }
3781 unlock:
3782         pte_unmap_unlock(vmf->pte, vmf->ptl);
3783         return 0;
3784 }
3785
3786 /*
3787  * By the time we get here, we already hold the mm semaphore
3788  *
3789  * The mmap_sem may have been released depending on flags and our
3790  * return value.  See filemap_fault() and __lock_page_or_retry().
3791  */
3792 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3793                 unsigned long address, unsigned int flags)
3794 {
3795         struct vm_fault vmf = {
3796                 .vma = vma,
3797                 .address = address & PAGE_MASK,
3798                 .flags = flags,
3799                 .pgoff = linear_page_index(vma, address),
3800                 .gfp_mask = __get_fault_gfp_mask(vma),
3801         };
3802         unsigned int dirty = flags & FAULT_FLAG_WRITE;
3803         struct mm_struct *mm = vma->vm_mm;
3804         pgd_t *pgd;
3805         p4d_t *p4d;
3806         vm_fault_t ret;
3807
3808         pgd = pgd_offset(mm, address);
3809         p4d = p4d_alloc(mm, pgd, address);
3810         if (!p4d)
3811                 return VM_FAULT_OOM;
3812
3813         vmf.pud = pud_alloc(mm, p4d, address);
3814         if (!vmf.pud)
3815                 return VM_FAULT_OOM;
3816         if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3817                 ret = create_huge_pud(&vmf);
3818                 if (!(ret & VM_FAULT_FALLBACK))
3819                         return ret;
3820         } else {
3821                 pud_t orig_pud = *vmf.pud;
3822
3823                 barrier();
3824                 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3825
3826                         /* NUMA case for anonymous PUDs would go here */
3827
3828                         if (dirty && !pud_write(orig_pud)) {
3829                                 ret = wp_huge_pud(&vmf, orig_pud);
3830                                 if (!(ret & VM_FAULT_FALLBACK))
3831                                         return ret;
3832                         } else {
3833                                 huge_pud_set_accessed(&vmf, orig_pud);
3834                                 return 0;
3835                         }
3836                 }
3837         }
3838
3839         vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3840         if (!vmf.pmd)
3841                 return VM_FAULT_OOM;
3842         if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3843                 ret = create_huge_pmd(&vmf);
3844                 if (!(ret & VM_FAULT_FALLBACK))
3845                         return ret;
3846         } else {
3847                 pmd_t orig_pmd = *vmf.pmd;
3848
3849                 barrier();
3850                 if (unlikely(is_swap_pmd(orig_pmd))) {
3851                         VM_BUG_ON(thp_migration_supported() &&
3852                                           !is_pmd_migration_entry(orig_pmd));
3853                         if (is_pmd_migration_entry(orig_pmd))
3854                                 pmd_migration_entry_wait(mm, vmf.pmd);
3855                         return 0;
3856                 }
3857                 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3858                         if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3859                                 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3860
3861                         if (dirty && !pmd_write(orig_pmd)) {
3862                                 ret = wp_huge_pmd(&vmf, orig_pmd);
3863                                 if (!(ret & VM_FAULT_FALLBACK))
3864                                         return ret;
3865                         } else {
3866                                 huge_pmd_set_accessed(&vmf, orig_pmd);
3867                                 return 0;
3868                         }
3869                 }
3870         }
3871
3872         return handle_pte_fault(&vmf);
3873 }
3874
3875 /*
3876  * By the time we get here, we already hold the mm semaphore
3877  *
3878  * The mmap_sem may have been released depending on flags and our
3879  * return value.  See filemap_fault() and __lock_page_or_retry().
3880  */
3881 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3882                 unsigned int flags)
3883 {
3884         vm_fault_t ret;
3885
3886         __set_current_state(TASK_RUNNING);
3887
3888         count_vm_event(PGFAULT);
3889         count_memcg_event_mm(vma->vm_mm, PGFAULT);
3890
3891         /* do counter updates before entering really critical section. */
3892         check_sync_rss_stat(current);
3893
3894         if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3895                                             flags & FAULT_FLAG_INSTRUCTION,
3896                                             flags & FAULT_FLAG_REMOTE))
3897                 return VM_FAULT_SIGSEGV;
3898
3899         /*
3900          * Enable the memcg OOM handling for faults triggered in user
3901          * space.  Kernel faults are handled more gracefully.
3902          */
3903         if (flags & FAULT_FLAG_USER)
3904                 mem_cgroup_enter_user_fault();
3905
3906         if (unlikely(is_vm_hugetlb_page(vma)))
3907                 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3908         else
3909                 ret = __handle_mm_fault(vma, address, flags);
3910
3911         if (flags & FAULT_FLAG_USER) {
3912                 mem_cgroup_exit_user_fault();
3913                 /*
3914                  * The task may have entered a memcg OOM situation but
3915                  * if the allocation error was handled gracefully (no
3916                  * VM_FAULT_OOM), there is no need to kill anything.
3917                  * Just clean up the OOM state peacefully.
3918                  */
3919                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3920                         mem_cgroup_oom_synchronize(false);
3921         }
3922
3923         return ret;
3924 }
3925 EXPORT_SYMBOL_GPL(handle_mm_fault);
3926
3927 #ifndef __PAGETABLE_P4D_FOLDED
3928 /*
3929  * Allocate p4d page table.
3930  * We've already handled the fast-path in-line.
3931  */
3932 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3933 {
3934         p4d_t *new = p4d_alloc_one(mm, address);
3935         if (!new)
3936                 return -ENOMEM;
3937
3938         smp_wmb(); /* See comment in __pte_alloc */
3939
3940         spin_lock(&mm->page_table_lock);
3941         if (pgd_present(*pgd))          /* Another has populated it */
3942                 p4d_free(mm, new);
3943         else
3944                 pgd_populate(mm, pgd, new);
3945         spin_unlock(&mm->page_table_lock);
3946         return 0;
3947 }
3948 #endif /* __PAGETABLE_P4D_FOLDED */
3949
3950 #ifndef __PAGETABLE_PUD_FOLDED
3951 /*
3952  * Allocate page upper directory.
3953  * We've already handled the fast-path in-line.
3954  */
3955 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
3956 {
3957         pud_t *new = pud_alloc_one(mm, address);
3958         if (!new)
3959                 return -ENOMEM;
3960
3961         smp_wmb(); /* See comment in __pte_alloc */
3962
3963         spin_lock(&mm->page_table_lock);
3964 #ifndef __ARCH_HAS_5LEVEL_HACK
3965         if (!p4d_present(*p4d)) {
3966                 mm_inc_nr_puds(mm);
3967                 p4d_populate(mm, p4d, new);
3968         } else  /* Another has populated it */
3969                 pud_free(mm, new);
3970 #else
3971         if (!pgd_present(*p4d)) {
3972                 mm_inc_nr_puds(mm);
3973                 pgd_populate(mm, p4d, new);
3974         } else  /* Another has populated it */
3975                 pud_free(mm, new);
3976 #endif /* __ARCH_HAS_5LEVEL_HACK */
3977         spin_unlock(&mm->page_table_lock);
3978         return 0;
3979 }
3980 #endif /* __PAGETABLE_PUD_FOLDED */
3981
3982 #ifndef __PAGETABLE_PMD_FOLDED
3983 /*
3984  * Allocate page middle directory.
3985  * We've already handled the fast-path in-line.
3986  */
3987 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3988 {
3989         spinlock_t *ptl;
3990         pmd_t *new = pmd_alloc_one(mm, address);
3991         if (!new)
3992                 return -ENOMEM;
3993
3994         smp_wmb(); /* See comment in __pte_alloc */
3995
3996         ptl = pud_lock(mm, pud);
3997 #ifndef __ARCH_HAS_4LEVEL_HACK
3998         if (!pud_present(*pud)) {
3999                 mm_inc_nr_pmds(mm);
4000                 pud_populate(mm, pud, new);
4001         } else  /* Another has populated it */
4002                 pmd_free(mm, new);
4003 #else
4004         if (!pgd_present(*pud)) {
4005                 mm_inc_nr_pmds(mm);
4006                 pgd_populate(mm, pud, new);
4007         } else /* Another has populated it */
4008                 pmd_free(mm, new);
4009 #endif /* __ARCH_HAS_4LEVEL_HACK */
4010         spin_unlock(ptl);
4011         return 0;
4012 }
4013 #endif /* __PAGETABLE_PMD_FOLDED */
4014
4015 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4016                             unsigned long *start, unsigned long *end,
4017                             pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4018 {
4019         pgd_t *pgd;
4020         p4d_t *p4d;
4021         pud_t *pud;
4022         pmd_t *pmd;
4023         pte_t *ptep;
4024
4025         pgd = pgd_offset(mm, address);
4026         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4027                 goto out;
4028
4029         p4d = p4d_offset(pgd, address);
4030         if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4031                 goto out;
4032
4033         pud = pud_offset(p4d, address);
4034         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4035                 goto out;
4036
4037         pmd = pmd_offset(pud, address);
4038         VM_BUG_ON(pmd_trans_huge(*pmd));
4039
4040         if (pmd_huge(*pmd)) {
4041                 if (!pmdpp)
4042                         goto out;
4043
4044                 if (start && end) {
4045                         *start = address & PMD_MASK;
4046                         *end = *start + PMD_SIZE;
4047                         mmu_notifier_invalidate_range_start(mm, *start, *end);
4048                 }
4049                 *ptlp = pmd_lock(mm, pmd);
4050                 if (pmd_huge(*pmd)) {
4051                         *pmdpp = pmd;
4052                         return 0;
4053                 }
4054                 spin_unlock(*ptlp);
4055                 if (start && end)
4056                         mmu_notifier_invalidate_range_end(mm, *start, *end);
4057         }
4058
4059         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4060                 goto out;
4061
4062         if (start && end) {
4063                 *start = address & PAGE_MASK;
4064                 *end = *start + PAGE_SIZE;
4065                 mmu_notifier_invalidate_range_start(mm, *start, *end);
4066         }
4067         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4068         if (!pte_present(*ptep))
4069                 goto unlock;
4070         *ptepp = ptep;
4071         return 0;
4072 unlock:
4073         pte_unmap_unlock(ptep, *ptlp);
4074         if (start && end)
4075                 mmu_notifier_invalidate_range_end(mm, *start, *end);
4076 out:
4077         return -EINVAL;
4078 }
4079
4080 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4081                              pte_t **ptepp, spinlock_t **ptlp)
4082 {
4083         int res;
4084
4085         /* (void) is needed to make gcc happy */
4086         (void) __cond_lock(*ptlp,
4087                            !(res = __follow_pte_pmd(mm, address, NULL, NULL,
4088                                                     ptepp, NULL, ptlp)));
4089         return res;
4090 }
4091